Systems and methods for virtual and augmented reality

Abstract

The description relates the feature matching. Our approach establishes pointwise correspondences between challenging image pairs. It takes off-the-shelf local features as input and uses an attentional graph neural network to solve an assignment optimization problem. The deep middle-end matcher acts as a middle-end and handles partial point visibility and occlusion elegantly, producing a partial assignment matrix.

Description

BACKGROUND OF THE INVENTION

1). Field of the Invention

This invention is related to connected mobile computing systems, methods, and configurations, and more specifically to mobile computing systems, methods, and configurations featuring at least one wearable component which may be utilized for virtual and/or augmented reality operation.

2). Discussion of Related Art

It is desirable that mixed reality, or augmented reality, near-eye displays be lightweight, low-cost, have a small form-factor, have a wide virtual image field of view, and be as transparent as possible. In addition, it is desirable to have configurations that present virtual image information in multiple focal planes (for example, two or more) in order to be practical for a wide variety of use-cases without exceeding an acceptable allowance for vergence-accommodation mismatch. Referring to FIG. 8, an augmented reality system is illustrated featuring a head-worn viewing component (2), a hand-held controller component (4), and an interconnected auxiliary computing or controller component (6) which may be configured to be worn as a belt pack or the like on the user. Each of these components may be operatively coupled (10, 12, 14, 16, 17, 18) to each other and to other connected resources (8) such as cloud computing or cloud storage resources via wired or wireless communication configurations, such as those specified by IEEE 802.11, Bluetooth®, and other connectivity standards and configurations. As described, for example, in U.S. patent application Ser. Nos. 14/555,585, 14/690,401, 14/331,218, 15/481,255, 62/627,155, 62/518,539, 16/229,532, 16/155,564, 15/413,284, 16/020,541, 62,702,322, 62/206,765, 15,597,694, 16/221,065, 15/968,673, 62/682,788, and 62/899,678 each of which is incorporated by reference herein in its entirety, various aspects of such components are described, such as various embodiments of the two depicted optical elements (20) through which the user may see the world around them along with visual components which may be produced by the associated system components, for an augmented reality experience. As illustrated in FIG. 8, such a system may also comprise various sensors configured to provide information pertaining to the environment around the user, including but not limited to various camera type sensors (such as monochrome, color/RGB, and/or thermal imaging components) (22, 24, 26), depth camera sensors (28), and/or sound sensors (30) such as microphones. There is a need for compact and persistently connected wearable computing systems and assemblies such as those described herein, which may be utilized to provide a user with the perception of rich augmented reality experiences.

SUMMARY OF THE INVENTION

This document describes certain aspects of what may be termed “the deep middle-end matcher”, a neural network configured to match two sets of local features by jointly finding correspondences and rejecting non-matchable points. Such a neural network configuration may be utilized in association with spatial computing resources such as those illustrated in FIG. 8, including but not limited to the camera and processing resources that comprise such spatial computing systems. Within the deep middle-end matcher type of configuration, assignments may be estimated by solving an optimal transport problem, whose costs are predicted by a graph neural network. We describe a flexible context aggregation mechanism based on attention, which enables the deep middle-end matcher configuration to reason about the underlying 3D scene and feature assignments jointly. Compared to traditional, hand-designed heuristics, our technique learns priors over geometric transformations and regularities of the 3D world through end-to-end training from images to correspondences. The deep middle-end matcher outperforms other learned approaches and sets a new state-of-the-art on the task of pose estimation in challenging real-world indoor and outdoor environments. These methods and configurations match in real-time on a modern graphical processing unit (“GPU”), and can be readily integrated into modern structure-from-motion (“SfM”) or simultaneous localization and mapping (“SLAM”) systems, all of which may be incorporated into systems such as that illustrated in FIG. 8.

The invention provides a computer system including a computer-readable medium, a processor connected to the computer-readable medium and a set of instructions on the computer-readable medium. The set of instructions may include a deep middle-end matcher architecture that may include an attentional graph neural network having a keypoint encoder to map keypoint positions p and their visual descriptors d into a single vector, and alternating self- and cross-attention layers that, based on the vector, repeated L times to create representations f; and an optimal matching layer that creates an M by N score matrix from the representations f and finds an optimal partial assignment based on the M by N score matrix.

The computer system may further include that in the keypoint encoder, an initial representation ⁽⁰⁾x_i for each keypoint i combines visual appearance and location, with the respective keypoint position embedded into a high-dimensional vector with a Multilayer Perceptron (MLP) as follows:⁽⁰⁾x_i=d_i+MLP(p_i)

The computer system may further include the keypoint encoder allows the attentional graph neural network to reason about appearance and position jointly.

The computer system may further include in the keypoint encoder includes a multiplex graph neural network having a single complete graph with nodes that are the keypoints of two images.

The computer system may further include the graph is a multiplex graph that has two types of undirected edges, namely intra-image edges (self edges; E_self) that connect keypoints i to all other keypoints within the same image and inter-image edges (cross edges, E_cross) that connect keypoints i to all keypoints in the other image and uses a message passing formulation to propagate information along both types of edges, such that the resulting multiplex graph neural network starts with a high-dimensional state for each node and computes at each layer an updated representation by simultaneously aggregating messages across all given edges for all nodes.

The computer system may further include if ^(l)x^A_i is the intermediate representation for element i in image A at layer l, the message m_E→i is the result of the aggregation from all keypoints {j:(i, j)∈E}, where E∈{E_self, E_cross}, and a residual message passing update for all i in A is:^(l+1)x_i^A=^(l)x_i^A+MLP ([^(l)x_i^A∥m_ε→i]).where [.∥.] denotes concatenation.

The computer system may further include a fixed number of layers L with different parameters are chained and alternatively aggregate along the self and cross edges such that, starting from l=1, E=E_self if l is odd and E=E_cross if l is even.

The computer system may further include the alternating self- and cross-attention layers are computed with an attention mechanism computes the message m_E→i and performs the aggregation, wherein the self edges are based on self-attention and the cross edges are based on cross-attention, wherein, for a representation of i, a query q_i, retrieves values v_j of some elements based on their attributes, the keys k_j, and the message is computed as weighted average of the values:

$m_{ℰ\to i}=\underset{j:\left(i,j\right)\in ℰ}{\Sigma }{\alpha }_{\mathrm{ij}}{v}_j.$

The computer system may further include an attention mask α_ij is the Softmax over the key-query similarities:αij=Softmax_j(q_i^Tk_j).

The computer system may further include the respective key, query, and value are computed as linear projections of deep features of the graph neural network, with a query keypoint i being in an image Q and all source keypoints are in image S, (Q, S)∈{A, B}², in the equation:

$q_i=W_1{i}_Q^{}+b_1\left[\begin{array}{c}{k}_j\\ v_j\end{array}\right]=\left[\begin{array}{c}{W}_2\\ W_3\end{array}\right]{ }_{}{i}_S^{}+\left[\begin{array}{c}{b}_2\\ b_3\end{array}\right].$

The computer system may further include final matching descriptors of the alternating self- and cross-attention layers are linear projections:

$f_i^A=W^{\left(L\right)}{x}_i^A+b\forall i\in A,$

The computer system may further include the optimal matching layer expresses a pairwise score for a set as the similarity of matching descriptors:

$S_{i,j}=<f_i^A,f_j^B>\forall \left(i,j\right)\in A\times B,$ where <., .> is the inner product. As opposed to learned visual descriptors, the matching descriptors are not normalized, and their magnitude can change per feature and during training to reflect the prediction confidence.

The computer system may further include the optimal matching layer, for occlusion and visibility suppresses occluded keypoints and augments each set of keypoints with a dustbin score so that unmatched keypoints are explicitly assigned to dustbin scores.

The computer system may further include the score S is augmented to S by appending a new row and column, the point-to-bin and bin-to-bin scores, filled with a single learnable parameter:

${\stackrel{\_}{S}}_{i,N+1}={\stackrel{\_}{S}}_{M+1,j}={\stackrel{\_}{S}}_{M+1,N+1}=\alpha \in ℝ.$

The computer system may further include the optimal matching layer finds the optimal partial assignment based on the M by N score matrix using the Sinkhorn algorithm for T iterations.

The computer system may further include after T iterations, we the optimal matching layer drops the dustbin scores and recovers P=P^T1_M≤1_N, whereP1_N≤1_M and P^T1_M≤1_N.is the original assignment andP1_N+1=a and P^T1_M+1=b. Is the assignment with the dustbin scored augmented.

The invention also provides a computer-implemented method system that may include mapping, with a keypoint encoder of an attentional graph neural network of a deep middle-end matcher architecture, keypoint positions p and their visual descriptors d into a single vector; and executing, with alternating self- and cross-attention layers of an attentional graph neural network of the deep middle-end matcher architecture, based on the vector, for L repeated times, to create representations f, and executing an optimal matching layer, of the attentional graph neural network of the deep middle-end matcher architecture, to create an M by N score matrix from the representations f and finding an optimal partial assignment based on the M by N score matrix.

BRIEF DESCRIPTION OF THE DRAWINGS

The invention is further described by way of example with reference to the accompanying drawings, wherein:

FIG. 1 is a representative sketch that illustrates feature matching with a deep middle-end matcher;

FIG. 2 shows correspondences estimated by the deep middle-end matcher for two difficult indoor image pairs;

FIG. 3 is a representative sketch that shows how we formulate the deep middle-end matcher to solve an optimization problem;

FIG. 4 are images that show masks as rays;

FIG. 5 are graphs that show indoor and outdoor pose estimation;

FIG. 6 shows qualitative image matches;

FIG. 7 shows visualizing attention in self- and cross-attention masks at various layers and heads; and

FIG. 8 shows an augmented reality system.

DETAILED DESCRIPTION OF THE INVENTION

Finding correspondences between points in images is a vital step for computer vision tasks dealing with 3D reconstruction or visual localization, such as Simultaneous Localization and Mapping (SLAM) and Structure-from-Motion (SfM). These estimate the 3D structure and camera poses from such correspondences after matching local features, a process known as data association. Factors such as large viewpoint change, occlusion, blur, and lack of texture make 2D-to-2D data association particularly challenging.

In this description, we present a new way of thinking about the feature matching problem. Instead of learning better task-agnostic local features followed by simple matching heuristics and tricks, we propose to learn the matching process from pre-existing local features using a novel neural architecture called the deep middle-end matcher (DMEM). In the context of SLAM that typically decomposes the problem into the visual feature detection front-end and the bundle adjustment or pose estimation back-end, our network lies directly in the middle—the deep middle-end matcher is a learnable middle-end. FIG. 1 illustrates feature matching with the deep middle-end matcher. Our approach establishes pointwise correspondences between challenging image pairs. It takes off-the-shelf local features as input and uses an attentional graph neural network to solve an assignment optimization problem. The deep middle-end matcher acts as a middle-end and handles partial point visibility and occlusion elegantly, producing a partial assignment matrix.

In this work, learning feature matching is viewed as finding the partial assignment between two sets of local features. We revisit the classical graph-based strategy of matching by solving a linear assignment problem, which, when relaxed to an optimal transport problem, can be solved differentiably [See references 50, 9, 31 below]. The cost function of this optimization is predicted by a Graph Neural Network (GNN). Inspired by the success of the Transformer [see reference 48 below], it uses self- (intra-image) and cross- (inter-image) attention to leverage both spatial relationships of the keypoints and their visual appearance. This formulation enforces the assignment structure of the predictions while enabling the cost to learn complex priors, elegantly handling occlusion and non-repeatable keypoints. Our method is trained end-to-end from images to correspondences—we learn priors for pose estimation from a large annotated dataset, enabling the deep middle-end matcher to reason about the 3D scene and the assignment. Our work can be applied to a variety of multiple-view geometry problems that require high-quality feature correspondences.

We show the superiority of the deep middle-end matcher compared to both handcrafted matchers and learned inlier classifiers. FIG. 2 shows correspondences estimated by the deep middle-end matcher for two difficult indoor image pairs. The deep middle-end matcher successfully estimates an accurate pose while other learned or handcrafted methods fail (correct correspondences in green). The proposed method brings the most substantial improvements when combined with SuperPoint [see reference 14 below], a deep front-end, thereby advancing the state-of-the-art on the tasks of homography estimation and indoor and outdoor pose estimation, and paving the way towards deep SLAM.

2. Related Work

Local feature matching is generally performed by i) detecting interest point, ii) computing visual descriptors, iii) matching these with a Nearest Neighbor (NN) search, iv) filtering incorrect matches, and finally v) estimating a geometric transformation. The classical pipeline developed in the 2000s is often based on SIFT [see reference 25 below], filters matches with Lowe's ratio test [see reference 25 below], the cross-check, and heuristics like neighborhood consensus [see references 46, 8, 5, 40 below], and finds a transformation with a robust solver like RANSAC [see references 17, 35 below].

Recent works on deep learning for matching often focus on learning better sparse detectors and local descriptors [see references 14, 15, 29, 37, 54 below] from data using Convolutional Neural Networks (CNNs). To improve their discriminativeness, some works explicitly look at a wider context using regional features [see reference 26 below] or log-polar patches [see reference 16 below]. Other approaches learn to filter matches by classifying them into inliers and outliers [see references 27, 36, 6, 56 below]. These operate on sets of matches, still estimated by NN search, and thus ignore the assignment structure and discard visual information. Works that actually learn to match have so far focused on dense matching [see reference 38 below] or 3D point clouds [see reference 52 below], and still exhibit such limitations. In contrast, our learnable middle-end simultaneously performs context aggregation, matching, and filtering in a single end-to-end architecture.

Graph matching problems are usually formulated as quadratic assignment problems, which are NP-hard, requiring expensive, complex, and thus impractical solvers [see reference 24 below]. For local features, the computer vision literature of the 2000s [see references 4, 21, 45 below] uses handcrafted costs with many heuristics, making it complex and fragile. Caetano et al. [see reference 7 below] learn the cost of the optimization for a simpler linear assignment, but only use a shallow model, while our deep middle-end matcher learns a flexible cost using a neural network. Related to graph matching is the problem of optimal transport [see reference 50 below]—it is a generalized linear assignment with an efficient yet simple approximate solution, the Sinkhorn algorithm [see references 43, 9, 31 below].

Deep Learning for sets such as point clouds aims at designing permutation equivariant or invariant functions by aggregating information across elements. Some works treat all of them equally, through global pooling [see references 55, 32, 11 below] or instance normalization [see references 47, 27, 26 below], while others focus on a local neighborhood in coordinate or feature space [see references 33, 53 below]. Attention [see references 48, 51, 49, 20 below] can perform both global and data-dependent local aggregation by focusing on specific elements and attributes, and is thus more flexible. Our work uses the fact that it can be seen as a particular instance of a Message Passing Graph Neural Network [see references 18, 3 below] on a complete graph. By applying attention to multi-edge, or multiplex, graphs, similar to [see references 22, 57 below], the deep middle-end matcher can learn complex reasoning about the two sets of local features.

3. The Deep Middle-End Matcher Architecture

Motivation: In the image matching problem, some regularities of the world could be leveraged: the 3D world is largely smooth and sometimes planar, all correspondences for a given image pair derive from a single epipolar transform if the scene is static, and some poses are more likely than others. In addition, 2D keypoints are usually projections of salient 3D points, like corners or blobs, thus correspondences across images must adhere to certain physical constraints: i) a keypoint can have at most a single correspondence in the other image; and ii) some keypoints will be unmatched due to occlusion and failure of the detector. An effective model for feature matching should aim at finding all correspondences between reprojections of the same 3D points and identifying keypoints that have no matches. FIG. 3 shows how we formulate the deep middle-end matcher as solving an optimization problem, whose cost is predicted by a deep neural network. The deep middle-end matcher is made up of two major components: the attentional graph neural network (Section 3a), and the optimal matching layer (Section 3b). The first component uses a keypoint encoder to map keypoint positions p and their visual descriptors d into a single vector, and then uses alternating self- and cross-attention layers (repeated L times) to create more powerful representations f. The optimal matching layer creates an M by N score matrix, augments it with dustbins, then finds the optimal partial assignment using the Sinkhorn algorithm (for T iterations). This alleviates the need for domain expertise and heuristics—we learn relevant priors directly from the data.

Formulation: Consider two images A and B, each with a set of keypoint positions p and associated visual descriptors d—we refer to them jointly (p, d) as the local features. Keypoints consist of x and y image coordinates as well as a detection confidence c, p_i:=(x, y, c)_i. Visual descriptors d_i∈R^D can be those extracted by a CNN like SuperPoint or traditional descriptors like SIFT. Images A and B have M and N local features and their sets of keypoint indices are A:{1, . . . , M} and B:={1, . . . , N}, respectively.

Partial Assignment: Constraints i) and ii) mean that correspondences derive from a partial assignment between the two sets of keypoints. For the integration into downstream tasks and better interpretability, each possible correspondence should have a confidence value. We consequently define a partial soft assignment matrix P∈[0, 1]^M×N as:P1_N≤1_M and P^T1_M≤1_N.  (1)

Our goal is the following: design a neural network that predicts the assignment P from two sets of local features.

3.1. Attentional Graph Neural Network

The first major block of the deep middle-end matcher (see Section 3a) is the attentional graph neural network whose job is the following: given initial local features, compute f_i∈R^D, the matching descriptors, by letting the features communicate with each other. Long-range feature communication is vital for robust matching and requires aggregation of information from within an image as well as across an image pair.

Intuitively, the distinctive information about a given keypoint depends on its visual appearance and its location, but also its spatial and visual relationship with respect to other co-visible keypoints, e.g. neighboring or salient ones. On the other hand, knowledge of keypoints in the second image can help to resolve ambiguities by comparing candidate matches or estimating the relative photometric or geometric transformation from global and unambiguous clues.

When asked to match a given ambiguous keypoint, humans look back-and-forth at both images: they sift for tentative matching keypoints, examine each of them, and look for contextual clues that help to disambiguate the true match from other self-similarities. This hints at an iterative process that can focus its attention on specific locations.

Keypoint Encoder: The initial representation ⁽⁰⁾x_i for each keypoint i combines visual appearance and location. We embed the keypoint position into a high-dimensional vector with a Multilayer Perceptron (MLP) as follows:⁽⁰⁾x_i=d_i+MLP(p_i).  (2)

The encoder allows the network to reason about both appearance and position jointly (this is especially powerful with an attention mechanism) and is an instance of the “positional encoder” introduced in the Transformer [see reference 48 below].

Multiplex Graph Neural Network: We consider a single complete graph whose nodes are the keypoints of both images. The graph has two types of undirected edges—it is a multiplex graph. Intra-image edges, or selfedges, E_self, connect keypoints i to all other keypoints within the same image. Inter-image edges, or cross edges, E_cross, connect keypoints i to all keypoints in the other image. We use the message passing formulation [see references 18, 3 below] to propagate information along both types of edges. The resulting Multiplex Graph Neural Network starts with a high-dimensional state for each node and computes at each layer an updated representation by simultaneously aggregating messages across all given edges for all nodes.

Let ^(l)x^A_i be the intermediate representation for element i in image A at layer l. The message m_E→i is the result of the aggregation from all keypoints {j:(i, j)∈E}, where E∈{E_self, E_cross}. The residual message passing update for all i in A is:^(l+1)x_i^A=^(l)x_i^A+MLP([^(l)x_i^A∥m_∈→i]).  (3)where [.∥.] denotes concatenation. A similar update can be simultaneously performed for all keypoints in image B. A fixed number of layers L with different parameters are chained and alternatively aggregate along the self and cross edges. As such, starting from l=1, E=E_self if l is odd and E=E_cross if l is even.

Attentional Aggregation: An attention mechanism computes the message m_E→i and performs the aggregation. Self edges are based on self-attention [see reference 48 below] and cross edges are based on cross-attention. Akin to database retrieval, a representation of i, the query q_i, retrieves the values v_j of some elements based on their attributes, the keys k_j. We compute the message as weighted average of the values:

$\begin{array}{cc}{m}_{ℰ\to i}=\underset{j:\left(i,j\right)\in ℰ}{\Sigma }{\alpha }_{\mathrm{ij}}{v}_j,& \left(4\right)\end{array}$ where the attention mask α_ij is the Softmax over the key-query similarities:

${\alpha }_{\mathrm{ij}}-{\mathrm{Softmax}}_j\left(q_i^{\top }{k}_j\right).$

The key, query, and value are computed as linear projections of deep features of the graph neural network. Considering that query keypoint i is in the image Q and all source keypoints are in image S, (Q, S)∈{A, B}², we can write:

$\begin{array}{cc}{q}_i=W_1{i}_Q^{}+b_1\left[\begin{array}{c}{k}_j\\ v_j\end{array}\right]=\left[\begin{array}{c}{W}_2\\ W_3\end{array}\right]{ }_{}{i}_S^{}+\left[\begin{array}{c}{b}_2\\ b_3\end{array}\right].& \left(5\right)\end{array}$

Each layer l has its own projection parameters, and they are shared for all keypoints of both images. In practice, we improve the expressivity with multi-headed attention [see reference 48 below].

Our formulation provides maximum flexibility as the network can learn to focus on a subset of key-points based on specific attributes. In FIG. 4, masks aij are shown as rays. Attentional aggregation builds a dynamic graph between keypoints. Self-attention (top) can attend anywhere in the same image, e.g., distinctive locations, and is thus not restricted to nearby locations. Cross-attention (bottom) attends to locations in the other image, such as potential matches that have a similar local appearance. The deep middle-end matcher can retrieve, or attend based on both appearance and keypoint location as they are encoded in the representation x_i. This includes attending to a nearby keypoint and retrieving the relative positions of similar or salient keypoints. This enables representations of the geometric transformation and the assignment. The final matching descriptors are linear projections:

$\begin{array}{cc}{f}_i^A=W^{\left(L\right)}{x}_i^A+b\forall i\in A,& \left(6\right)\end{array}$ and similarly for keypoints in B.

3.2. Optimal Matching Layer

The second major block of the deep middle-end matcher (see Section 3b) is the optimal matching layer, which produces a partial assignment matrix. As in the standard graph matching formulation, the assignment P can be obtained by computing a score matrix S∈R^M×N for all possible matches and maximizing the total score Σ_i,j S_i,j P_i,j under constraints in Equation 1. This is equivalent to solving a linear assignment problem.

Score Prediction: Building a separate representation for all (M+1)×(N+1) potential matches would be prohibitive. We instead express the pairwise score as the similarity of matching descriptors:

$\begin{array}{cc}{S}_{i,j}=<f_i^A,f_j^B>\forall \left(i,j\right)\in A\times B,& \left(7\right)\end{array}$ where <., .> is the inner product. As opposed to learned visual descriptors, the matching descriptors are not normalized, and their magnitude can change per feature and during training to reflect the prediction confidence.

Occlusion and Visibility: To let the network suppress occluded keypoints, we augment each set with a dustbin so that unmatched keypoints are explicitly assigned to it. This technique is common in graph matching, and dustbins have also been used by SuperPoint [see reference 14 below] to account for image cells that might not have a detection. We augment the score S to S by appending a new row and column, the point-to-bin and bin-to-bin scores, filled with a single learnable parameter:

$\begin{array}{cc}{\stackrel{\_}{S}}_{i,N+1}={\stackrel{\_}{S}}_{M+1,j}={\stackrel{\_}{S}}_{M+1,N+1}=\alpha \in ℝ.& \left(8\right)\end{array}$

While keypoints in A will be assigned to a single keypoint in B or the dustbin, each dustbin has as many matches as there are keypoints in the other set: N, M for dustbins in A, B respectively. We denote as a=[1_M^T N]^T and b=[1_N^T M]^T the number of expected matches for each keypoint and dustbin in A and B. The augmented assignment P now has the constraints:P1_N+1=a and P^T1_M+1=b.  (9)

Sinkhorn Algorithm: The solution of the above optimization problem corresponds to the optimal transport [see reference 31 below] between discrete distributions a and b with score S. It can be approximately solved with the Sinkhorn algorithm [see references 43, 9 below], a differentiable version of the Hungarian algorithm [see reference 28 below], classically used for bipartite matching. It solves a regularized transport problem, naturally resulting in a soft assignment. This normalization amounts to iteratively performing alternating Softmax along rows and columns, and is thus easily parallelized on GPU. After T iterations, we drop the dustbins and recover P=P_1:M,1:N.

3.3. Loss

By design, both the graph neural network and the optimal matching layer are differentiable—this enables backpropagation from matches to visual descriptors. The deep middle-end matcher is trained in a supervised manner from ground truth matches M={(i, j)}⊂A×B. These are estimated from ground truth relative transformations—using poses and depth maps or homographies. This also lets us label some keypoints I⊆A and J⊆B as unmatched if they do not have any reprojection in their vicinity. Given the labels, we minimize the negative log-likelihood of the assignment P:

$\begin{array}{cc}\begin{array}{c}\mathrm{Loss}=-\underset{\left(i,j\right)\in ℳ}{\Sigma }\log {\stackrel{\_}{P}}_{i,j}\\ -\underset{i\in ℐ}{\Sigma }\log {\stackrel{\_}{P}}_{i,N+1}-\underset{j\in 𝒥}{\Sigma }\log {\stackrel{\_}{P}}_{M+1,j}.\end{array}& \left(10\right)\end{array}$

This supervision aims at simultaneously maximizing the precision and the recall of the matching.

3.4. Comparisons to Related Work

The deep middle-end matcher vs. inlier classifiers [see references 27, 56 below]: the deep middle-end matcher benefits from a strong inductive bias by being entirely permutation equivariant with respect to both images and local features. It additionally embeds the commonly-used mutual check constraint directly into the training: any match with probability P_i,j greater than 0.5 is necessarily mutually consistent.

The deep middle-end matcher vs. Instance Normalization[see reference 47 below]: Attention, as used by the deep middle-end matcher, is a more flexible and powerful context aggregation mechanism than instance normalization, which treats all keypoints equally and is used by previous work on feature matching [see references 27, 56, 26 below].

The deep middle-end matcher vs. ContextDesc[see reference 26 below]: the deep middle-end matcher can jointly reason about appearance and position while ContextDesc processes them separately. Additionally, ContextDesc is a front-end that additionally requires a larger regional extractor, and a loss for keypoints scoring. The deep middle-end matcher only needs local features, learned or handcrafted, and can thus be a simple drop-in replacement of existing matchers.

The deep middle-end matcher vs. Transformer[see reference 48 below]: the deep middle-end matcher borrows the self-attention from the Transformer, but embeds it into a graph neural network, and additionally introduces the cross-attention, which is symmetric. This simplifies the architecture and results in better feature reuse across layers.

4. Implementation Details

The deep middle-end matcher can be combined with any local feature detector and descriptor but works particularly well with SuperPoint [see reference 14 below], which produces repeatable and sparse keypoints—enabling very efficient matching. Visual descriptors are bilinearly sampled from the semi-dense feature map, which is differentiable. Both local feature extraction and subsequent “gluing” are directly performed on the GPU. At test time, in order to extract matches from the soft assignment, one can use the confidence threshold to retain some, or simply use all of them and their confidence in a subsequent step, such as weighted pose estimation.

Architecture details: All intermediate representations (key, query value, descriptors) have the same dimension D=256 as the SuperPoint descriptors. We use L=9 layers of alternating multi-head self- and cross-attentions with 4 heads each, and perform T=100 Sinkhorn iterations—in log-space for numerical stability. The model is implemented in PyTorch [see reference 30 below] and runs in real-time on a GPU: a forward pass takes on average 150 ms (7 FPS).

Training details: To allow for data augmentation, SuperPoint detect and describe steps are performed on-the-fly as batches during training. A number of random keypoints are further added for efficient batching and increased robustness. More details are provided in Appendix A.

5. Experiments

5.1. Homography Estimation

We perform a large-scale homography estimation experiment using real images and synthetic homographies with both robust (RANSAC) and non-robust (DLT) estimators.

Dataset: We generate image pairs by sampling random homographies and applying random photometric distortions to real images, following a recipe similar to [see references 12, 14, 37, 36 below]. The underlying images come from the set of 1M distractor images in the Oxford and Paris dataset [see reference 34 below], split into training, validation, and test sets.

Baselines: We compare the deep middle-end matcher against several matchers applied to SuperPoint local features—the Nearest Neighbor (NN) matcher and various outlier rejectors: the mutual check (or cross-check), PointCN [see reference 27 below], and Order-Aware Network (OANet) [see reference 56 below]. All learned methods, including the deep middle-end matcher, are trained on ground-truth correspondences, found by projecting keypoints from one image to the other. We generate homographies and photometric distortions on-the-fly—an image pair is never seen twice during training.

Metrics: Match precision (P) and recall (R) are computed from the ground truth correspondences. Homography estimation is performed with both RANSAC and the Direct Linear Transformation [see reference 19 below] (DLT), which has a direct least-squares solution. We compute the mean reprojection error of the four corners of the image and report the area under the cumulative error curve (AUC) up to a value of 10 pixels.

Results: the deep middle-end matcher is sufficiently expressive to master homographies, achieving 98% recall and high precision. Table 1 shows Homography estimation for the deep middle-end matcher, DLT and RANSAC. The deep middle-end matcher recovers al-most all possible matches while suppressing most outliers. Because the deep middle-end matcher correspondences are high-quality, the Direct Linear Transform (DLT), a least-squares based solution with no robustness mechanism, outperforms RANSAC. The estimated correspondences are so good that a robust estimator is not required—the deep middle-end matcher works even better with DLT than RANSAC. Outlier rejection methods like PointCN and OANet cannot predict more correct matches than the NN matcher itself, overly relying on the initial descriptors.

TABLE 1
		Homography estimation
Local		AUC
features	Matcher	RANSAC	DLT	P	R
SuperPoint	NN	39.47	0.00	21.7	65.4
	NN + mutual	42.45	0.24	43.8	56.5
	NN + PointCN	43.02	45.40	76.2	64.2
	NN + OANet	44.31	49.85	80.8	64.5
	DMEM	53.67	65.85	90.7	98.3

5.2. Indoor Pose Estimation

Indoor image matching is very challenging due to the lack of texture, the abundance of self-similarities, the complex 3D geometry of scenes, and large viewpoint changes. As we show in the following, the deep middle-end matcher can effectively learn priors to overcome these challenges.

Dataset: We use ScanNet [see reference 10 below], a large-scale indoor dataset composed of monocular sequences with ground truth poses and depth images as well as well-defined training, validation, and test splits corresponding to different scenes. Past works select training and evaluation pairs based on time difference [see references 29, 13 below] or SfM co-visibility [see references 27, 56, 6 below], usually computed using SIFT. We argue that this limits the difficulty of the pairs, and instead select these based on an overlap score computed for all possible image pairs in a given sequence using only ground truth poses and depth. This results in significantly wider-baseline pairs, which corresponds to the current frontier for real-world indoor image matching. Discarding pairs with too small or too large overlap, we obtain 230M training pairs and sample 1500 test pairs. More details are provided in Appendix A.

Metrics: As in previous work [see references 27, 56, 6 below], we report the AUC of the pose error at the thresholds (5°, 10°, 20°), where the pose error is the maximum of the angular errors in rotation and translation. Relative pose is obtained from essential matrix estimation with RANSAC. We also report the match precision and the matching score [see references 14, 54 below], where a match is deemed correct based on its epipolar distance.

Baselines: We evaluate the deep middle-end matcher and various baseline matchers using both root-normalized SIFT [see references 25, 2 below] and SuperPoint [see reference 14 below] features. The deep middle-end matcher is trained with correspondences and unmatched keypoints derived from ground-truth poses and depth. All baselines are based on the Nearest Neighbor (NN) matcher and potentially an outlier rejection method. In the “Handcrafted” category, we consider the simple cross-check (mutual), ratio test [see reference 25 below], descriptor distance threshold, and the more complex GMS [see reference 5 below]. Methods in the “Learned” category are PointCN [see reference 27 below], and its follow-ups OANet [see reference 56 below] and NG-RANSAC [see reference 6 below]. We retrain PointCN and OANet on ScanNet for both SuperPoint and SIFT with the classification loss using the above-defined correctness criterion and their respective regression losses. For NG-RANSAC, we use the original trained model. We do not include any graph matching methods as they are orders of magnitude too slow for the number of keypoints that we consider. We report other local features as reference: ORB [see reference 39 below] with GMS, D2-Net [see reference 15 below], and ContextDesc [see reference 26 below] using the publicly available trained models.

Results: the deep middle-end matcher enables significantly higher pose accuracy compared to both handcrafted and learned matchers. Table 2 shows wide-baseline indoor pose estimation on Scan-Net. We report the AUC of the pose error, the matching score (MS) and precision (P), all in Pose estimation AUC percents. The deep middle-end matcher outperforms all handcrafted and learned matchers when applied to both SIFT and SuperPoint. These benefits are substantial when applied to both SIFT and SuperPoint. FIG. 5 shows indoor and outdoor pose estimation. The deep middle-end matcher significantly improves the pose accuracy over OANet, a state-of-the-art outlier rejection neural network. It has a significantly higher precision than other learned matchers, demonstrating its higher representation power. It also produces a larger number of correct matches—up to 10 times more than the ratio test when applied to SIFT, because it operates on the full set of possible matches, rather than the limited set of nearest neighbors. SuperPoint and the deep middle-end matcher together achieve state-of-the-art results on indoor pose estimation. They complement well each other since repeatable keypoints make it possible to estimate a larger number of correct matches even in very challenging situations—see FIG. 2.

TABLE 2
		Pose estimation
Local		AUC
features	Matcher	@5°	@10°	@20°	P	MS
ORB	NN + GMS	5.21	13.65	25.36	72.0	5.7
D2-Net	NN + mutual	5.25	14.53	27.96	46.7	12.0
ContextDesc	NN + ratio test	6.64	15.01	25.75	51.2	9.2
SIFT	NN + ratio test	5.83	13.06	22.47	40.3	1.0
	NN + OANet	5.77	13.17	23.93	38.0	4.3
	NN + NG-RANSAC	6.19	13.80	23.73	61.9	0.7
	DMEM	6.71	15.70	28.67	74.2	9.8
SuperPoint	NN + mutual	9.43	21.53	36.40	50.4	18.8
	NN + distance +	9.82	22.42	36.83	63.9	14.6
	mutual
	NN + GMS	8.39	18.96	31.56	50.3	19.0
	NN + PointCN	11.40	25.47	41.41	71.8	25.5
	NN + OANet	11.76	26.90	43.85	74.0	25.7
	DMEM	16.16	33.81	51.84	84.4	31.5

In FIG. 6 shows qualitative image matches. We compare the deep middle-end matcher to the Nearest Neighbor (NN) matcher with two outlier rejectors, handcrafted and learned, in three environments. The deep middle-end matcher consistently estimates more correct matches (green lines) and fewer mismatches (red lines), coping with repeated texture, large viewpoint, and illumination changes.

5.3. Outdoor Pose Estimation

As outdoor image sequences present their own set of challenges (e.g., lighting changes and occlusion), we train and evaluate the deep middle-end matcher for pose estimation in an outdoor setting. We use the same evaluation metrics and baseline methods as in the indoor pose estimation task.

Dataset: We evaluate on the PhotoTourism dataset, which is part of the CVPR'19 Image Matching Challenge [see reference 1 below]. It is a subset of the YFCC100M dataset [see reference 44 below] and has ground truth poses and sparse 3D models obtained from an off-the-shelf SfM tool [see references 29, 41, 42 below]. For training, we use the MegaDepth dataset [see reference 23 below], which also has clean depth maps computed with multi-view stereo. Scenes that are in the PhotoTourism test set are removed from the training set.

Results: Table 3 shows outdoor pose estimation on the PhotoTourism dataset. Matching SuperPoint and SIFT features with the deep middle-end matcher results in significantly higher pose accuracy (AUC), precision (P), and matching score (MS) than with handcrafted or other learned methods. The deep middle-end matcher outperforms all baselines, at all relative pose thresholds, when applied to both SuperPoint and SIFT. Most notably, the precision of the resulting matching is very high (84.9%), reinforcing the analogy that the deep middle-end matcher “glues” together local features.

TABLE 3
		Pose estimation
Local		AUC
features	Matcher	@5°	@10°	@20°	P	MS
ContextDesc	NN + ratio test	20.16	31.65	44.05	56.2	3.3
SIFT	NN + ratio test	15.19	24.72	35.30	43.4	1.7
	NN + NG-RANSAC	15.61	25.28	35.87	64.4	1.9
	NN + OANet	17.87	27.85	39.43	53.2	3.5
	DMEM	23.03	36.51	50.07	74.0	7.3
SuperPoint	NN + mutual	9.80	18.99	30.88	22.5	4.9
	NN + GMS	13.96	24.58	36.53	47.1	4.7
	NN + OANet	21.03	34.08	46.88	52.4	8.4
	DMEM	34.18	50.32	64.16	84.9	11.1

5.4. Understanding the Deep Middle-End Matcher

Ablation Study: To evaluate our design decisions, we repeated the indoor ScanNet experiments, but this time focusing on different the deep middle-end matcher variants. Table 4 shows Ablation of the deep middle-end matcher on ScanNet with Super-Point local features. that all the deep middle-end matcher blocks are useful and bring substantial performance gains. Differences with respect to the full model are shown. While the optimal matching layer alone improves over the baseline Nearest Neighbor matcher, the GNN ex-plains the majority of the gains brought by the deep middle-end matcher. Both cross-attention and positional encoding are critical for strong gluing, and a deeper net further improves precision.

TABLE 4
		Pose	Match	Matching
Matcher		AUC@20°	precision	score
NN +		36.40	59.4	18.8
mutual
DMEM	No Graph Neural Net	38.56	66.0	17.2
	No cross-attention	42.57	74.0	25.3
	No positional encoding	47.12	75.8	26.6
	Smaller (3 layers)	46.93	79.9	30.0
	Full (9 layers)	51.84	84.4	31.5

Visualizing Attention: An understanding of the proposed technique would not be complete without an attempt to visualize the deep middle-end matcher's attention patterns throughout matching. The extensive diversity of self- and cross-attention patterns is shown in FIG. 7 and reflects the complexity of the learned behavior. FIG. 7 shows visualizing attention: self- and cross-attention masks α_ij at various layers and heads. The deep middle-end matcher learns a diversity of patterns and can focus on global or local context, self-similarities, distinctive features, and match candidates.

6. Conclusion

In this disclosure, we described what we refer to as “the deep middle-end matcher”—an attentional graph neural network inspired by the Trans-former's success in NLP—for local feature matching. We believe that the data association component of the 3D reconstruction pipeline has not received adequate attention from the research community, and powerful learning-based middle-ends are our solution. The deep middle-end matcher boosts the receptive field of local features and downplays features whose correspondences are missing, effectively performing the roles of both ContextDesc and inlier classification. Importantly, the inner-workings of the deep middle-end matcher are learned entirely from real-world data. Our results on 2D-to-2D feature matching show significant improvement over the existing state-of-the-art.

Our description herein makes a strong case for the use of learnable middle-ends in the feature matching pipeline as a modern, deep learning-based, alternative to hand-designed heuristics. Some of our future work will focus on evaluating the deep middle-end matcher inside a complete 3D reconstruction pipeline.

Various example embodiments of the invention are described herein. Reference is made to these examples in a non-limiting sense. They are provided to illustrate more broadly applicable aspects of the invention. Various changes may be made to the invention described and equivalents may be substituted without departing from the true spirit and scope of the invention. In addition, many modifications may be made to adapt a particular situation, material, composition of matter, process, process act(s) or step(s) to the objective(s), spirit or scope of the present invention. Further, as will be appreciated by those with skill in the art that each of the individual variations described and illustrated herein has discrete components and features which may be readily separated from or combined with the features of any of the other several embodiments without departing from the scope or spirit of the present inventions. All such modifications are intended to be within the scope of claims associated with this disclosure.

The invention includes methods that may be performed using the subject devices. The methods may comprise the act of providing such a suitable device. Such provision may be performed by the end user. In other words, the “providing” act merely requires the end user obtain, access, approach, position, set-up, activate, power-up or otherwise act to provide the requisite device in the subject method. Methods recited herein may be carried out in any order of the recited events which is logically possible, as well as in the recited order of events.

Example aspects of the invention, together with details regarding material selection and manufacture have been set forth above. As for other details of the present invention, these may be appreciated in connection with the above-referenced patents and publications as well as generally known or appreciated by those with skill in the art. The same may hold true with respect to method-based aspects of the invention in terms of additional acts as commonly or logically employed.

In addition, though the invention has been described in reference to several examples optionally incorporating various features, the invention is not to be limited to that which is described or indicated as contemplated with respect to each variation of the invention. Various changes may be made to the invention described and equivalents (whether recited herein or not included for the sake of some brevity) may be substituted without departing from the true spirit and scope of the invention. In addition, where a range of values is provided, it is understood that every intervening value, between the upper and lower limit of that range and any other stated or intervening value in that stated range, is encompassed within the invention.

Also, it is contemplated that any optional feature of the inventive variations described may be set forth and claimed independently, or in combination with any one or more of the features described herein. Reference to a singular item, includes the possibility that there are plural of the same items present. More specifically, as used herein and in claims associated hereto, the singular forms “a,” “an,” “said,” and “the” include plural referents unless the specifically stated otherwise. In other words, use of the articles allow for “at least one” of the subject item in the description above as well as claims associated with this disclosure. It is further noted that such claims may be drafted to exclude any optional element. As such, this statement is intended to serve as antecedent basis for use of such exclusive terminology as “solely,” “only” and the like in connection with the recitation of claim elements, or use of a “negative” limitation.

Without the use of such exclusive terminology, the term “comprising” in claims associated with this disclosure shall allow for the inclusion of any additional element—irrespective of whether a given number of elements are enumerated in such claims, or the addition of a feature could be regarded as transforming the nature of an element set forth in such claims. Except as specifically defined herein, all technical and scientific terms used herein are to be given as broad a commonly understood meaning as possible while maintaining claim validity.

The breadth of the present invention is not to be limited to the examples provided and/or the subject specification, but rather only by the scope of claim language associated with this disclosure.

7. Appendix A—Further Experimental Details

Homography Estimation:

The test set contains 1024 pairs of 640×480 images. Homographies are generated by applying random perspective, scaling, rotation, and translation to the original full-sized images, to avoid bordering artifacts. We evaluate with the 512 top-scoring keypoints detected by SuperPoint with a Non-Maximum Suppression (NMS) radius of 4 pixels. Correspondences are deemed correct if they have a reprojection error lower than 3 pixels. When estimating the homography with RANSAC, we use the OpenCV function findHomography with 3000 iterations and an inlier threshold of 3 pixels.

Indoor Pose Estimation:

The overlap score between two images A and B is the average ratio of pixels in A that are visible in B (and vice versa), after accounting for missing depth values and occlusion (by checking for consistency in the depth using relative errors). We train and evaluate with an overlap range of 0.4 to 0.8. For training, we sample at each epoch 200 pairs per scene, similarly as in [15]. The test set is generated by subsampling the sequences by 15 and subsequently sampling 15 pairs for each of the 300 sequences. We resize all ScanNet images and depth maps to VGA 640×480. We detect up to 1024 SuperPoint keypoints (using the publicly available trained model with NMS radius of 4) and 2048 SIFT keypoints (using OpenCV's implementation). When computing the precision and matching score, we use an epipolar threshold of 5.10e-4. Poses are computed by first estimating the essential matrix with OpenCV's findEssentialMat and RANSAC with an inlier threshold of 1 pixel divided by the average focal length, followed by recoverPose. In contrast with previous works [28, 59, 6], we compute a more accurate AUC using explicit integration rather than coarse histograms.

Outdoor Pose Estimation:

For training on Megadepth, the overlap score is the ratio of triangulated keypoints that are visible in the two images, as in [15]. We sample pairs with an overlap score in [0.1, 0.7] at each epoch. For the evaluation on the PhotoTourism dataset, we use all 11 scenes and the overlap score computed by Ono [30], with a selection range of [0.1, 0.4]. Images are resized so that their longest edge is smaller than 1600 pixels. We detect 2048 keypoints for both SIFT and SuperPoint (with an NMS radius of 3). Other evaluation parameters are identical to the ones used in the indoor evaluation.

Training of the Deep Middle-End Matcher:

For training on homography/indoor/outdoor data, we use the Adam optimizer with an initial constant learning rate of 10e-4 for the first 200k/100k/50k iterations, followed by an exponential decay of 0.999998/0.999992/0.999992 until iteration 900k. When using SuperPoint features, we employ batches with 32/64/16 image pairs and a fixed number of 512/400/1024 keypoints per image. When using SIFT features we use 1024 keypoints and 24 pairs. Because of the limited number of training scenes, the outdoor model is initialized with the homography model. Prior to the keypoint encoding, keypoints are normalized by the largest edge of the image.

Ground truth correspondences M and unmatched sets I and J are generated by first computing the M×N re-projection matrix between all detected keypoints using the ground truth homography or pose and depth map. Correspondences are cells that have a reprojection error that is a minimum along both rows and columns, and that is lower than a given threshold: 3, 5, and 3 pixels for homographies, indoor, and outdoor matching respectively. For homographies, unmatched keypoints are simply the ones that do not appear in M. For indoor and outdoor matching, because of errors in the depth and the pose, unmatched keypoints must additionally have a minimum reprojection error larger than 15 and 5 pixels, respectively. This allows to ignore labels for keypoints whose correspondences are ambiguous, while still providing some supervision through the Sinkhorn normalization.

8. References

Each of the following references is incorporated by reference herein in its entirety, and referenced in the above description:

• [1] Phototourism Challenge, CVPR 2019 Image Matching Workshop. https://image matching-workshop. github.io. Accessed Nov. 8, 2019. 7
• [2] Relja Arandjelović and Andrew Zisserman. Three things everyone should know to improve object retrieval. In CVPR, 2012. 6
• [3] Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks. arXiv:1806.01261, 2018. 2, 3
• [4] Alexander C Berg, Tamara L Berg, and Jitendra Malik. Shape matching and object recognition using low distortion correspondences. In CVPR, 2005. 2
• [5] JiaWang Bian, Wen-Yan Lin, Yasuyuki Matsushita, Sai-Kit Yeung, Tan-Dat Nguyen, and Ming-Ming Cheng. GMS: Grid-based motion statistics for fast, ultra-robust feature correspondence. In CVPR, 2017. 2, 6
• [6] Eric Brachmann and Carsten Rother. Neural-Guided RANSAC: Learning Where to Sample Model Hypotheses. In ICCV, 2019. 2, 6
• Tibério S Caetano, Julian J McAuley, Li Cheng, Quoc V Le, and Alex J Smola. Learning graph matching. IEEE TPAMI, 31(6):1048-1058, 2009. 2
• [8] Jan Cech, Jiri Matas, and Michal Perdoch. Efficient sequential correspondence selection by cosegmentation. IEEE TPAMI, 32(9):1568-1581, 2010. 2
• [9] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NIPS, 2013. 1, 2, 4
• [10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nieβner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 6
• [11] Haowen Deng, Tolga Birdal, and Slobodan Ilic. Ppfnet: Global context aware local features for robust 3d point matching. In CVPR, 2018. 2
• [12] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. In RSS Work-shop: Limits and Potentials of Deep Learning in Robotics, 2016. 5
• [13] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Self-improving visual odometry. arXiv:1812.03245, 2018. 6
• [14] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. SuperPoint: Self-supervised interest point detection and description. In CVPR Workshop on Deep Learningfor Visual SLAM, 2018. 2, 4, 5, 6
• [15] Mihai Dusmanu, Ignacio Rocco, Tomas Pajdla, Marc Pollefeys, Josef Sivic, Akihiko Torii, and Torsten Sattler. D2-net: A trainable cnn for joint detection and description of local features. In CVPR, 2019. 2, 6
• [16] Patrick Ebel, Anastasiia Mishchuk, Kwang Moo Yi, Pascal Fua, and Eduard Trulls. Beyond cartesian representations for local descriptors. In ICCV, 2019. 2
• [17] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2
• [18] Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In ICML, 2017. 2, 3
• [19] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 6
• [20] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A frame-work for attention-based permutation-invariant neural networks. In ICML, 2019. 2
• [21] Marius Leordeanu and Martial Hebert. A spectral technique for correspondence problems using pairwise constraints. In ICCV, 2005. 2
• [22] Yujia Li, Chenjie Gu, Thomas Dullien, Oriol Vinyals, and Pushmeet Kohli. Graph matching networks for learning the similarity of graph structured objects. In ICML, 2019. 2
• [23] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In CVPR, 2018. 7
• [24] Eliane Maria Loiola, Nair Maria Maia de Abreu, Paulo Oswaldo Boaventura-Netto, Peter Hahn, and Tania Querido. A survey for the quadratic assignment problem. European journal of operational research, 176(2):657-690, 2007. 2
• [25] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 2, 6
• [26] Zixin Luo, Tianwei Shen, Lei Zhou, Jiahui Zhang, Yao Yao, Shiwei Li, Tian Fang, and Long Quan. Contextdesc: Local descriptor augmentation with cross-modality context. In CVPR, 2019. 2, 5, 6
• [27] Kwang Moo Yi, Eduard Trulls, Yuki Ono, Vincent Lepetit, Mathieu Salzmann, and Pascal Fua. Learning to find good correspondences. In CVPR, 2018. 2, 5, 6
• [28] James Munkres. Algorithms for the assignment and transportation problems. Journal of the society for industrial and applied mathematics, 5(1):32-38, 1957. 4
• [29] Yuki Ono, Eduard Trulls, Pascal Fua, and Kwang Moo Yi. LF-Net: Learning local features from images. In NeurIPS, 2018. 2, 6, 7
• [30] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Al-ban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NIPS Workshops, 2017. 5
• [31] Gabriel Peyr6, Marco Cuturi, et al. Computational optimal transport. Foundations and Trends in Machine Learning, 11(5-6):355-607, 2019. 1, 2, 4
• [32] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2
• [33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NIPS, 2017. 2
• [34] Filip Radenovic, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018. 5, 9
• [35] Rahul Raguram, Jan-Michael Frahm, and Marc Pollefeys. A comparative analysis of ransac techniques leading to adaptive real-time random sample consensus. In ECCV, 2008. 2
• [36] René Ranftl and Vladlen Koltun. Deep fundamental matrix estimation. In ECCV, 2018. 2, 5
• [37] Jerome Revaud, Philippe Weinzaepfel, Ceésar DeSouza, Noe Pion, Gabriela Csurka, Yohann Cabon, and Martin Humenberger. R2D2: Repeatable and reliable detector and descriptor. In NeurIPS, 2019. 2, 5
• [38] Ignacio Rocco, Mircea Cimpoi, Relja Arandjelovic, Akihiko Torii, Tomas Pajdla, and Josef Sivic. Neighbourhood consensus networks. In NeurIPS, 2018. 2
• [39] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary R Bradski. Orb: An efficient alternative to sift or surf. In ICCV, 2011. 6
• [40] Torsten Sattler, Bastian Leibe, and Leif Kobbelt. Scramsac: Improving ransac's efficiency with a spatial consistency fil-ter. In ICCV, 2009. 2
• [41] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 7
• [42] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for un-structured multi-view stereo. In ECCV, 2016. 7
• [43] Richard Sinkhorn and Paul Knopp. Concerning nonnegative matrices and doubly stochastic matrices. Pacific Journal of Mathematics, 1967. 2, 4
• [44] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 7
• [45] Lorenzo Torresani, Vladimir Kolmogorov, and Carsten Rother. Feature correspondence via graph matching: Models and global optimization. In ECCV, 2008. 2
• [46] Tinne Tuytelaars and Luc J Van Gool. Wide baseline stereo matching based on local, affinely invariant regions. In BMVC, 2000. 2
• [47] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization. arXiv:1607.08022, 2016. 2, 5
• [48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko-reit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 1, 2, 3, 4, 5
• [49] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro LiŠó, and Yoshua Bengio. Graph attention networks. In ICLR, 2018. 2
• [50] Cédric Villani. Optimal transport: old and new, volume338. Springer Science & Business Media, 2008. 1, 2
• [51] Xiaolong Wang, Ross Girshick, Abhinav Gupta, and Kaiming He. Non-local neural networks. In CVPR, 2018. 2
• [52] Yue Wang and Justin M Solomon. Deep Closest Point: Learning representations for point cloud registration. In ICCV, 2019. 2
• [53] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E. Sarma, Michael M. Bronstein, and Justin M. Solomon. Dynamic Graph CNN for learning on point clouds. ACM Transactions on Graphics, 2019. 2
• [54] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In ECCV, 2016. 2, 6
• [55] Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnabas Poczos, Ruslan R Salakhutdinov, and Alexander J Smola. Deep sets. In NIPS, 2017. 2
• [56] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Hon-gen Liao. Learning two-view correspondences and geometry using order-aware network. In ICCV, 2019. 2, 5, 6
• [57] Li Zhang, Xiangtai Li, Anurag Arnab, Kuiyuan Yang, Yunhai Tong, and Philip HS Torr. Dual graph convolutional net-work for semantic segmentation. In BMVC, 2019.

Claims

1. A computer system comprising: a computer-readable medium;a processor connected to the computer-readable medium; anda set of instructions on the computer-readable medium, including:a deep middle-end matcher architecture that includes: an attentional graph neural network having: a keypoint encoder to map keypoint positions p and their visual descriptors d into a single vector; andalternating self-attention and cross-attention layers that, based on the vector, repeated L times to create representations f; andan optimal matching layer that creates an M by N score matrix from the representations f and finds an optimal partial assignment based on the M by N score matrix,where:p are keypoint positions,d are descriptors,L are a plurality of times,f are representations,M by N is a score matrix with a length M and a width N.

2. The computer system of claim 1, wherein, in the keypoint encoder, an initial representation (0)xi for each keypoint combines visual appearance and location, with the respective keypoint position embedded into a high-dimensional vector with a Multilayer Perceptron as follows: (0)xi=di+MLP(pi).where:i are keypoints,(0)xi is an initial representation for the keypoints,di are descriptors for the keypoints,pi are keypoint positions for the keypoints, andMLP is a Multilayer Perceptron for the keypoint positions.

3. The computer system of claim 2, wherein the keypoint encoder allows the attentional graph neural network to reason about appearance and position jointly.

4. The computer system of claim 1, wherein, in the keypoint encoder includes a multiplex graph neural network having a single complete graph with nodes that are the keypoints of two images.

5. The computer system of claim 4, wherein the graph is a multiplex graph that has two types of undirected edges, namely intra-image edges Eself that connect keypoints i to all other keypoints within the same image and inter-image edges Ecross that connect keypoints i to all keypoints in the other image and uses a message passing formulation to propagate information along both types of edges, such that the resulting multiplex graph neural network starts with a high-dimensional state for each node and computes at each layer an updated representation by simultaneously aggregating messages across all given edges for all nodes where:Eself are self edges, andEcross are cross edges.

6. The computer system of claim 5, wherein if (l)xAi is the intermediate representation for keypoint i in image A at layer l, the message mE→i is the result of the aggregation from all keypoints {j:(i, j)∈E}, where E∈{Eself, Ecross}, and a residual message passing update for all i in A is: (l+1)xiA=(l)xiA+MLP([(l)xiA∥mε→i]),where:[.∥.] denotes concatenation,i are keypoints,Eself are self edges,Ecross are cross edges,mE is a message, andMLP is a Multilayer Perceptron for the keypoint positions.

7. The computer system of claim 6, wherein a fixed number of layers with different parameters are chained and alternatively aggregate along the self and cross edges such that, starting from l=1, E=Eself if l is odd and E=Ecross if l is even.

8. The computer system of claim 6, wherein the alternating self- and cross-attention layers are computed with an attention mechanism computes the message mE→i and performs the aggregation, wherein the self edges are based on self-attention and the cross edges are based on cross-attention, wherein, for a representation of i, a query qi, retrieves values vj of some elements based on their attributes, the keys kj, and the message is computed as weighted average of the values:

9. The computer system of claim 8, wherein an attention mask αij is the Softmax over the key-query similarities: αij=Softmaxj(qiTkj).

10. The computer system of claim 1, wherein final matching descriptors of the alternating self- and cross-attention layers are linear projections:

11. The computer system of claim 1, wherein the optimal matching layer expresses a pairwise score for a set as the similarity of matching descriptors:

12. The computer system of claim 11, wherein the optimal matching layer, for occlusion and visibility suppresses occluded keypoints and augments each set of keypoints with a dustbin score so that unmatched keypoints are explicitly assigned to dustbin scores.

13. The computer system of claim 12, wherein the score S is augmented to S by appending a new row and column, the point-to-bin and bin-to-bin scores, filled with a single learnable parameter: Si,N+1=SM+1,j=SM+1,N+1=α∈R. where:N+1 and M+1 represent augmentations.

14. The computer system of claim 12, wherein the optimal matching layer finds the optimal partial assignment based on the M by N score matrix using the Sinkhorn algorithm for T iterations.

15. The computer system of claim 14, wherein after T iterations, the optimal matching layer drops the dustbin scores and recovers P=P−1:M,1:N, where P1N+1=a and PT1M≤1N.

16. A computer-implemented method system that includes: mapping, with a keypoint encoder of an attentional graph neural network of a deep middle-end matcher architecture, keypoint positions p and their visual descriptors d into a single vector; andexecuting, with alternating self- and cross-attention layers of an attentional graph neural network of the deep middle-end matcher architecture, based on the vector, for L repeated times, to create representations f; and executing an optimal matching layer, of the attentional graph neural network of the deep middle-end matcher architecture, to create an M by N score matrix from the representations f and finding an optimal partial assignment based on the M by N score matrix,where:p are keypoint positions,d are descriptors,L are a plurality of times,f are representations,M by N is a score matrix with a length M and a width N.