Systems and methods for training pose estimators in computer vision

Abstract

A data capture stage includes a frame at least partially surrounding a target object, a rotation device within the frame and configured to selectively rotate the target object, a plurality of cameras coupled to the frame and configured to capture images of the target object from different angles, a sensor coupled to the frame and configured to sense mapping data corresponding to the target object, and an augmentation data generator configured to control a rotation of the rotation device, to control operations of the plurality of cameras and the sensor, and to generate training data based on the images and the mapping data.

Description

FIELD

Aspects of embodiments of the present disclosure are generally related to training data generation systems and methods of using the same.

BACKGROUND

In many areas of automation, such as robotics, sensors are used to determine the physical relationship of objects in the real world. For example, robotic systems often use sensing systems to measure the locations of various physical objects in order to, for example, grasp an object that may arrive at a variety of orientations, reorient the object into a desired position, and connect the object to another object. The position and orientation of an object with respect to a reference coordinate system may be referred to as a “pose” and, in a three-dimensional coordinate system, generally includes six degrees of freedom—rotation around three axes and translation along the three axes. Recently, statistical models, such as machine learning models, are being utilized to control the operation of such automation systems.

Statistical models (e.g., machine learning models) are generally trained using large amounts of data. In the field of computer vision, the training data generally includes labeled images, which are used to train deep learning models, such as convolutional neural networks, to perform computer vision tasks such as image classification, instance segmentation, and pose estimation. However, manually collecting photographs of various scenes and labeling the photographs is time consuming and expensive. Some techniques for augmenting these training data sets include generating synthetic training data. For example, three-dimensional (3-D) computer graphics rendering engines (e.g., scanline rendering engines and ray tracing rendering engines) are capable of generating photorealistic two-dimensional (2-D) images of virtual environments of arrangements of 3D models of objects that can be used for training deep learning models.

However, synthetic data may not capture the full range of nuances and complexities that exist in the real world. As such, systems trained with synthetic data do not always perform well when presented with real-world data.

The above information disclosed in this Background section is only for enhancement of understanding of the present disclosure, and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art.

SUMMARY

Aspects of embodiments of the present disclosure relate to the generation of data for training machine learning models. In particular, aspects of embodiments of the present disclosure relate to generating a large quantity of real-world data for training machine learning models to perform computer vision tasks on input images that are captured based on visible light in a scene or on imaging modalities other than images of the intensity of visible light in a scene.

Aspects of embodiments of the present disclosure relate to systems and methods for generating and using visual datasets for training computer vision models including object pose detection models.

Aspects of embodiments of the present disclosure relate to particular techniques for generating labeled data based on the pose of a single object or the poses of multiple objects in a cluttered bin of those objects.

According to some embodiments of the present invention, there is provided a data capture stage including: a frame at least partially surrounding a target object; a rotation device within the frame and configured to selectively rotate the target object; a plurality of cameras coupled to the frame and configured to capture images of the target object from different angles; a sensor coupled to the frame and configured to sense mapping data corresponding to the target object; and an augmentation data generator configured to control a rotation of the rotation device, to control operations of the plurality of cameras and the sensor, and to generate training data based on the images and the mapping data.

In some embodiments, the data capture stage further includes a plurality of light sources coupled to the frame and configured to illuminate the target object with different colors and from different angles.

In some embodiments, the augmentation data generator is configured to control a color intensity of a light source of the plurality of light sources.

In some embodiments, a light source of the plurality of light sources is moveably coupled to the frame, and the augmentation data generator is further configured to control a position of the light source.

In some embodiments, the plurality of cameras includes: a color camera configured to capture red-green-blue (RGB) images of the target object; and a polarization camera configured to capture polarization images of the target object.

In some embodiments, the sensor includes at least one of an infrared sensor, an ultraviolet sensor, and a depth sensor, and the depth sensor includes at least one of a light detection and ranging (LIDAR) sensor and a stereo camera system.

In some embodiments, the plurality of cameras and the sensor are moveably coupled to the frame, and the augmentation data generator is further configured to control poses of the cameras and the sensor.

In some embodiments, the plurality of cameras and the sensor are angled to face the target object, and the rotation device is configured to rotate the target object with a 0.001 degree precision.

In some embodiments, the augmentation data generator is configured to: rotate the rotation device by a first angle; initiate capture of first images by the cameras and initiate sensing of first mapping data by the sensor; rotate the rotation device by a second angle; initiate capture of second images by the cameras and initiate sensing of second mapping data by the sensor; and generate the training data based on the first and second images and the first and second mapping data.

In some embodiments, the augmentation data generator is configured to: initiate capture of a first image of the target object and a second image of the target object by the plurality of cameras; receive a first pose of the target object corresponding to a first viewpoint of the first image; project the first pose onto the second image to generate a second pose of the target object corresponding to a second viewpoint of the second image; generate a second label associated with the second image, the second label including the second pose; and generating the training data based on the second image and the second label.

According to some embodiments of the present invention, there is provided a method of capturing training data in a data capture stage, the method including: initiating capture of a first image of a target object by a first camera of the data capture stage; initiating capture of a second image of the target object by a second camera of the data capture stage; receiving a first pose of the target object corresponding to a first viewpoint of the first image; projecting the first pose onto the second image to generate a second pose of the target object corresponding to a second viewpoint of the second image; generating a second label associated with the second image, the second label including the second pose; and generating the training data based on the second image and the second label.

In some embodiments, the first viewpoint is different from the second viewpoint, and the first image is a red-green-blue (RGB) image and the second image is an RGB image or a polarization image.

In some embodiments, the first and second poses include six-degree-of-freedom (6DoF) poses of the target object.

In some embodiments, the second label further includes at least one of a computer-aided design (CAD) model of the target object posed according to the second pose and key points corresponding to the CAD model.

In some embodiments, the method further includes: generating a first label associated with the first image, the first label including the first pose, wherein the generating the training data is further based on the first image and the first label.

In some embodiments, the method further includes: initiating rotation of a rotation device of the data capture stage on which the target object is placed; initiating capture of a third image of the target object, the third image corresponding to a viewpoint different from the first and second viewpoints; projecting the first pose onto the third image to generate a third pose of the target object corresponding to a third viewpoint of the third image; generating a third label associated with the third image, the third label including the third pose; and generating the training data further based on the third image and the third label.

In some embodiments, the method further includes: initiating a first lighting condition via a plurality of light sources of the data capture stage, wherein initiating the capture of the first image is in response to the initiating the first lighting condition; and initiating a second lighting condition via the plurality of light sources of the data capture stage, wherein initiating the capture of the second image is in response to the initiating the second lighting condition.

In some embodiments, the method further includes: initiating sensing of mapping data of the target object by a sensor of the data capture stage, wherein the mapping data includes one of a heat map corresponding to the target object, a depth map corresponding to the target object, and an ultraviolet map corresponding to the target object.

In some embodiments, the method further includes: projecting the first pose onto the mapping data to generate a third pose of the target object corresponding to a third viewpoint of the mapping data; generating a third label associated with the mapping data, the third label including the third pose; and generating the training data further based on the mapping data and the third label.

In some embodiments, the method further includes: providing the first and second images, a computer-aided design (CAD) model of the target object, and calibration information of the first and second cameras to a user device, wherein receiving the first pose of the target object includes receiving the first pose from the user device.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings, together with the specification, illustrate example embodiments of the present disclosure, and, together with the description, serve to explain the principles of the present disclosure.

FIG. 1 is a block diagram depicting a system for training a statistical model to perform computer vision tasks based on images in various modalities, where the training is performed using data generated in accordance with embodiments of the present disclosure.

FIG. 2 is a schematic block diagram of a computer vision system configured to use polarization imaging and that can generate training data based on polarization imaging data generated according to some embodiments of the present invention.

FIG. 3 is a schematic diagram of the data capture stage, according to some example embodiments of the present disclosure.

FIGS. 4A-4D illustrate images taken of an arrangement of objects from four different viewpoints of four cameras, according to some embodiments of the present disclosure.

FIGS. 5A-5D illustrate the images taken from four different viewpoints overlaid with initial pose estimates of an object's CAD model, according to some embodiments of the present disclosure.

FIGS. 6A-6B illustrate refined poses of an object across two viewpoints as overlays of key points, according to some embodiments of the present disclosure.

FIGS. 7A-7B illustrate the refined poses across the two viewpoints as overlays of a rendered CAD model, according to some embodiments of the present disclosure.

FIGS. 8A-8B is a flow diagram illustrating a process of capturing and augmenting training data, according to some example embodiments of the present disclosure.

DETAILED DESCRIPTION

The detailed description set forth below is intended as a description of example embodiments of a system and method for generating a large set of training data from a single object or an arrangement of the same objects, provided in accordance with the present disclosure, and is not intended to represent the only forms in which the present disclosure may be constructed or utilized. The description sets forth the features of the present disclosure in connection with the illustrated embodiments. It is to be understood, however, that the same or equivalent functions and structures may be accomplished by different embodiments that are also intended to be encompassed within the scope of the disclosure. As denoted elsewhere herein, like element numbers are intended to indicate like elements or features.

Automation systems that manipulate objects, such as tools, partially assembled products, components, and the like from a particular manufacturing environment, rely on pose estimations (i.e., six degree of freedom (6DoF) pose estimations) of the object based on some visual input. Training statistical models that perform 6DoF pose estimation often involves a great deal of training data. Sometimes this data is in the form of real images with ground truth pose estimates. However, oftentimes, the number of available labeled data is too small to properly train a pose estimator. Further, manually collecting different images of typical objects and scenes in the manufacturing environment and labeling these images based on their ground truth values is generally a time consuming and expensive task.

Accordingly, aspects of the present disclosure are directed to systems and methods for amplifying data, that is, by taking one real scene with an arrangement of identical or substantially identical objects (with labeled 6DoF poses), and then applying data augmentation techniques that can consistently be applied to the 6DoF labels. Since at least one data sample needs ground truth label, aspects of embodiments of the present disclosure provide a multi-view pose annotation technique for labeling a 6DoF pose from multiple viewpoints. In some embodiments, an augmentation data generator uses a scene with 6DoF labels to generate labeled variants of the scene, at different viewpoints, exposure conditions, and optical properties (such as color, polarimetry, etc).

Aspects of embodiments of the present disclosure relate to systems and methods for generating data for training machine learning models for performing computer vision tasks on images captured based on standard modalities such as color or monochrome cameras configured to capture images based on the intensity of visible light, as well as non-standard modalities such as polarization images captured based on polarized light (e.g., images captured with a polarizing filter or polarization filter in an optical path of the camera for capturing circularly and/or linearly polarized light), non-visible or invisible light (e.g., light in the infrared or ultraviolet (UV) ranges), and combinations thereof (e.g., polarized infrared light). However, embodiments of the present disclosure are not limited thereto and may be applied to other multi-spectral imaging techniques.

FIG. 1 is a block diagram depicting a system for training a statistical model to perform computer vision tasks based on images in various modalities, where the training is performed using data generated in accordance with embodiments of the present disclosure. As shown in FIG. 1, training data 5 is supplied to a model training system 7, which takes a model 30 (e.g., a pre-trained model or a model architecture with initialized weights) and uses the training data 5 to generate a trained model (or re-trained model) 32. The model 30 and the trained model 32 may be a statistical model such as a deep neural network (deep neural networks include convolutional neural networks). An augmentation data generator 220 from a data capture stage 100 according to embodiments of the present disclosure generates data 101, which may be included with the training data 5 for generating the trained model 32. The model training system 7 may apply an iterative process for updating the parameters of the model 30 to generate the trained model 32 in accordance with the supplied training data 5. The updating of the parameters of the model 30 may include, for example, applying gradient descent (and, in the case of a neural network, backpropagation) in accordance with a loss function measuring a difference between the labels and the output of the model in response to the training data. The model training system 7 and the augmentation data generator 220 may be implemented using one or more electronic circuits.

According to various embodiments of the present disclosure, the model training system 7 and/or the augmentation data generator 220 are implemented using one or more electronic circuits configured to perform various operations as described in more detail below. Types of electronic circuits may include a central processing unit (CPU), a graphics processing unit (GPU), an artificial intelligence (AI) accelerator (e.g., a vector processor, which may include vector arithmetic logic units configured efficiently perform operations common to neural networks, such dot products and softmax), a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a digital signal processor (DSP), or the like. For example, in some circumstances, aspects of embodiments of the present disclosure are implemented in program instructions that are stored in a non-volatile computer readable memory where, when executed by the electronic circuit (e.g., a CPU, a GPU, an AI accelerator, or combinations thereof), perform the operations described herein to capture and augment training data based on input from a plurality of cameras and sensors. The operations performed by the model training system 7 and the augmentation data generator 220 may be performed by a single electronic circuit (e.g., a single CPU, a single GPU, or the like) or may be allocated between multiple electronic circuits (e.g., multiple GPUs or a CPU in conjunction with a GPU). The multiple electronic circuits may be local to one another (e.g., located on a same die, located within a same package, or located within a same embedded device or computer system) and/or may be remote from one other (e.g., in communication over a network such as a local personal area network such as Bluetooth®, over a local area network such as a local wired and/or wireless network, and/or over wide area network such as the internet, such a case where some operations are performed locally and other operations are performed on a server hosted by a cloud computing service). One or more electronic circuits operating to implement the model training system 7 and the augmentation data generator 220 may be referred to herein as a computer or a computer system, which may include memory storing instructions that, when executed by the one or more electronic circuits, implement the systems and methods described herein.

FIG. 2 is a schematic block diagram of a computer vision system configured to use polarization imaging and that can generate training data based on polarization imaging data generated according to some embodiments of the present invention.

For context, FIG. 2 is a schematic depiction of a system in which a polarization camera images a scene and supplies polarization raw frames (also referred to as polarization images) to a computer vision system that includes a model that is trained to perform computer vision tasks based on polarization raw frames or polarization features computed based on polarization raw frames

A polarization camera 10 has a lens 12 with a field of view, where the lens 12 and the polarization camera 10 are oriented such that the field of view encompasses the scene 1. The lens 12 is configured to direct light (e.g., focus light) from the scene 1 onto a light sensitive medium such as an image sensor 14 (e.g., a complementary metal oxide semiconductor (CMOS) image sensor or charge-coupled device (CCD) image sensor).

The polarization camera 10 further includes a polarizer or polarizing filter or polarization mask 16 placed in the optical path between the scene 1 and the image sensor 14. According to some embodiments of the present disclosure, the polarizer or polarization mask 16 is configured to enable the polarization camera 10 to capture images of the scene 1 with the polarizer set at various specified angles (e.g., at 45° rotations or at 60° rotations or at non-uniformly spaced rotations).

As one example, FIG. 2 depicts an embodiment where the polarization mask 16 is a polarization mosaic aligned with the pixel grid of the image sensor 14 in a manner similar to a red-green-blue (RGB) color filter (e.g., a Bayer filter) of a color camera. In a manner similar to how a color filter mosaic filters incoming light based on wavelength such that each pixel in the image sensor 14 receives light in a particular portion of the spectrum (e.g., red, green, or blue) in accordance with the pattern of color filters of the mosaic, a polarization mask 16 using a polarization mosaic filters light based on linear polarization such that different pixels receive light at different angles of linear polarization (e.g., at 0°, 45°, 90°, and 135°, or at 0°, 60° degrees, and 120°). Accordingly, the polarization camera 10 using a polarization mask 16 such as that shown in FIG. 2 is capable of concurrently or simultaneously capturing light at four different linear polarizations.

While the above description relates to some possible implementations of a polarization camera using a polarization mosaic, embodiments of the present disclosure are not limited thereto and encompass other types of polarization cameras that are capable of capturing images at multiple different polarizations. For example, the polarization mask 16 may have fewer than or more than four different polarizations, or may have polarizations at different angles (e.g., at angles of polarization of: 0°, 60° degrees, and 120° or at angles of polarization of 0°, 30°, 60°, 90°, 120°, and 150°). As another example, the polarization mask 16 may be implemented using an electronically controlled polarization mask, such as an electro-optic modulator (e.g., may include a liquid crystal layer), where the polarization angles of the individual pixels of the mask may be independently controlled, such that different portions of the image sensor 14 receive light having different polarizations. As another example, the electro-optic modulator may be configured to transmit light of different linear polarizations when capturing different frames, e.g., so that the camera captures images with the entirety of the polarization mask set to, sequentially, to different linear polarizer angles (e.g., sequentially set to: 0 degrees; 45 degrees; 90 degrees; or 135 degrees). As another example, the polarization mask 16 may include a polarizing filter that rotates mechanically, such that different polarization raw frames are captured by the polarization camera 10 with the polarizing filter mechanically rotated with respect to the lens 12 to transmit light at different angles of polarization to image sensor 14.

A polarization camera may also refer to an array of multiple cameras having substantially parallel optical axes, such that each of the cameras captures images of a scene from substantially the same pose. The optical path of each camera of the array includes a polarizing filter, where the polarizing filters have different angles of polarization. For example, a two-by-two (2×2) array of four cameras may include one camera having a polarizing filter set at an angle of 0°, a second camera having a polarizing filter set at an angle of 45°, a third camera having a polarizing filter set at an angle of 90°, and a fourth camera having a polarizing filter set at an angle of 135°.

As a result, the polarization camera captures multiple input images 18 (or polarization raw frames) of the scene 1, where each of the polarization raw frames 18 corresponds to an image taken behind a polarization filter or polarizer at a different angle of polarization ϕ_pol (e.g., 0 degrees, 45 degrees, 90 degrees, or 135 degrees). Each of the polarization raw frames is captured from substantially the same pose with respect to the scene 1 (e.g., the images captured with the polarization filter at 0 degrees, 45 degrees, 90 degrees, or 135 degrees are all captured by a same polarization camera located at a same location and orientation), as opposed to capturing the polarization raw frames from disparate locations and orientations with respect to the scene. The polarization camera 10 may be configured to detect light in a variety of different portions of the electromagnetic spectrum, such as the human-visible portion of the electromagnetic spectrum, red, green, and blue portions of the human-visible spectrum, as well as invisible portions of the electromagnetic spectrum such as infrared and ultraviolet.

The augmentation data generator 220 of the data capture stage 100 utilizes the input images 18, in addition to other input data (e.g., RGB images, sensed map data, etc.), to generate training data 101 for a multitude of 2D and 3D vision tasks.

FIG. 3 is a schematic diagram of the data capture stage 200, according to some example embodiments of the present disclosure.

In some embodiments, the data capture stage 200 includes a frame 202, a rotation device (e.g., a turntable) 204, a plurality of cameras 206, a plurality of sensors 208, a plurality of light sources 210, and an augmentation data generator 220, which are used to collect training data based on a target object 300 from an arrangement 302 of target objects 300. The object 300 may be opaque, shiny metallic, transparent/semi-transparent, or the like.

The frame 202 may include a plurality of skeletal struts that create an enclosure. The frame may have a spherical shape, as shown in FIG. 3, or may have any suitable geometry, such as a cylinder, a cuboid, or the like. The frame may at least partially surround the rotation device 204 and the target object 300 placed upon it. The cameras 206, the sensors 208, and the light sources 210 are coupled to (e.g., moveably coupled to) the frame 202. In some examples, at least some of the cameras 206, the sensors 208, and the light sources 210 may be coupled to the frame 202 via motorized moving mechanisms (e.g., including rails and rollers) that can be controlled by a controller (e.g., stage controller) 222 of the augmentation data generator 220. Thus, in such examples, the controller 222 may adjust the position and orientation of at least some of cameras 206, the sensors 208, and the light sources 210 as desired. The data capture stage 200 may be modular such that cameras 20, sensors 208, and the light sources 210 can easily be added or removed.

In some examples, different light sources 210 may be capable of producing different color lights (e.g., red, green, and blue). Each of the light sources 210 may be moveably coupled to the frame 202 (e.g., coupled to the skeletal struts of the frame 202) and include one or more light emitting diodes (LEDs) with adjustable brightness/intensity. In some embodiments, the controller 222 is configured to control the on/off status, the brightness, and the position (e.g., vertical position) of the light sources 210 within the frame 202. As such, the data capture stage 200 is capable of creating many unique lighting condition that may be desired for illuminating the target object 300, such as a particular lighting condition suitable for measuring reflectance. In some examples, the data capture stage 200 is capable of reproducing lighting conditions that mimic those of a manufacturing environment for which the data capture stage 200 is generating training data.

The cameras 206 may include RGB cameras (also referred to as visible-light or color cameras) capable of capturing color images (e.g., RGB images) in the visible range of light and polarization cameras 10 that are capable of capturing polarized light, as described above. The cameras 206 may be connected to the frame 202 at different points around the target object 300 with overlapping fields of view. This allows the cameras 206 to capture images of the target object 300 from different but overlapping viewpoints.

In some embodiments, the various individual cameras 206 are registered with one another by determining their relative poses (or relative positions and orientations) by capturing multiple images of a calibration target, such as a checkerboard pattern, an ArUco target (see, e.g., Garrido-Jurado, Sergio, et al. “Automatic generation and detection of highly reliable fiducial markers under occlusion.” Pattern Recognition 47.6 (2014): 390-402), or a ChArUco target (see, e.g., An, Gwon Hwan, et al. “Charuco board-based omnidirectional camera calibration method.” Electronics 7.12 (2018): 421). In particular, the process of calibrating the targets may include computing intrinsic matrices characterizing the internal parameters of each camera 206 (e.g., matrices characterizing the focal length, image sensor format, and principal point of the camera) and extrinsic matrices characterizing the pose of each camera with respect to world coordinates (e.g., matrices for performing transformations between camera coordinate space and world or scene coordinate space). Different cameras 206 may have image sensors with different sensor formats (e.g., aspect ratios) and/or different resolutions without limitation, and the computed intrinsic and extrinsic parameters of the individual cameras enable the augmentation data generator 220 to map different portions of the different images from the various cameras 206 to a same coordinate space (where possible, such as where the fields of view overlap).

The sensors 208 may capture different parameters of a plenoptic function (such as polarization information, infrared and UV data, etc.). In some embodiments, the sensors 208 include at least one of an infrared sensor configured to capture the temperature variations over the target object 300, a depth sensor configured to capture depth information of the target object 300, and the like. According to some examples, the depth sensor may be a LIDAR (light detection and ranging) sensor that emits pulses of light and determines depth based on time-of-flight measurements, or may be a stereo camera configuration (e.g., a stereo polarization camera system) that computes depth estimates based on parallax shifts between stereo pairs of images (e.g., stereo pair of RGB or polarization images). While the disclosure herein provides examples of depth and infrared sensors, embodiments of the present disclosure are not limited thereto, and the data capture stage 200 may include any suitable type of sensor such as a sensor for capturing a UV texture maps, etc. The sensors 208 may be calibrated by using the same checkboard discussed above and/or any other suitable object (such as a cube or sphere) to estimate the extrinsic parameters of the sensors (see, e.g., Eung-su Kim, et al. “Extrinsic Calibration between Camera and LiDAR Sensors by Matching Multiple 3D Planes.” Sensors 2020, 20, 52).

According to some embodiments, the controller 222 controls the operations of the cameras 206 and sensors 208 and, for example, may be capable of syncing the capture time of the cameras 206.

The rotation device 204 includes a table top (e.g. a flat table top) that is configured to rotate about a vertical axis (e.g., about an axis orthogonal to the table top) with high precision (e.g., with a precision of 0.001 degrees). As the cameras 206 and the sensors 208 are substantially stationary during image capture and the sensing of mapping data, the rotatable rotation device 204 allows the target object to be imaged and sensed from almost any arbitrary angle.

According to some embodiments, in addition to calibrating the relative position and orientation of the cameras 206, the data capture stage 200 calibrates the cameras 206 for different rotations of the rotation device 204. With the calibration target on the tabletop, the cameras 206 capture a first set of images, then the controller 222 rotates the rotation device 204 by a set amount (e.g., 1 degree) and the cameras 206 capture a second set of images. The first and second set of images allow the augmentation data generator 220 to calculate the transformation matrix for the set rotation. Using this, the augmentation data generator 220 can generate the transformation matrix for any rotation. For example, it can be shown that:Transformation Matrix at X degrees=(Transformation Matrix at 1 degree)^X

The same transformation matrix may be utilized to calibrate the sensors for different angles.

As will be discussed further below, the augmentation data generator 220 utilizes the rotation transformations to automatically generate a plethora labeled training data from a single scene.

In some embodiments, the augmentation data generator 220 includes a pose estimator 224 configured to compute or estimate a pose of the target objects 300 based on information captured by the cameras 10 and, in some examples, based on the mapping data sensed by the sensors 208. However, embodiments of the present disclosure are not limited thereto, and the pose estimator 224 may exist within the user device 230, or may not be utilized.

In particular, a “pose” refers to the position and orientation of an object with respect to a reference coordinate system. For example, a reference coordinate system may be defined with a camera 206 at the origin, where the direction along the optical axis of the camera 206 (e.g., a direction through the center of its field of view) is defined as the z-axis of the coordinate system, and the x and y axes are defined to be perpendicular to one another and perpendicular to the z-axis. (Embodiments of the present disclosure are not limited to this particular coordinate system, and a person having ordinary skill in the art would understand that poses can be mathematically transformed to equivalent representations in different coordinate systems.)

Generally, in a three-dimensional coordinate system, an objects has six degrees of freedom—rotation around three axes (e.g., rotation around x-, y-, and z-axes) and translation along the three axes (e.g., translation along x-, y-, and z-axes). As such, an object's “pose” may be defined as its six degrees of freedom (6DoF) pose.

In some embodiments, it is assumed that a three-dimensional (3-D) model or computer aided design (CAD) model representing a canonical or ideal version of the target object 300 is available. For example, in some embodiments of the present disclosure, the target object 300 is an individual instance of manufactured components that have a substantially uniform appearance from one component to the next. Examples of such manufactured components include screws, bolts, nuts, connectors, and springs, as well as specialty parts such electronic circuit components (e.g., packaged integrated circuits, light emitting diodes, switches, resistors, and the like), laboratory supplies (e.g. test tubes, PCR tubes, bottles, caps, lids, pipette tips, sample plates, and the like), and manufactured parts (e.g., handles, switch caps, light bulbs, and the like). Accordingly, in these circumstances, a CAD model defining the ideal or canonical shape of the target object 300 may be used to define a coordinate system for the object (e.g., the coordinate system used in the representation of the CAD model).

In some examples, the pose estimator 224 utilizes at least one of the images (e.g., RGB images) of the target object 300 in the overlapping and different fields of view of the cameras 206 and the CAD model of the target object 300 to generate an initial pose estimate of the object 300. In embodiments in which the pose estimator 224 is within the augmentation data generator 220, this initial pose estimate is provided to user device 230. In other examples, the user device 230 houses the pose estimator 224 and utilizes the initial pose estimate. The user device 230 renders projections of the CAD model with the initial estimated pose onto the different images. The projection may be performed by using object-to-camera and camera-to-world transforms to rotate the CAD model to a particular pose and then render the 3D model onto a 2D image of the object 300. The device 230 allows a user to manually manipulate and refine the object pose to create a refined pose that is more accurate than the initial pose estimate generated by the pose estimator 224. This refined pose is used by the augmentation data generator 220 to create ground truth poses or labels for a plethora of images and data captured by the cameras 206 and sensors 208.

FIGS. 4A-4D illustrate images taken of an arrangement 302 of objects 300 from four different viewpoints of the cameras 206, according to some embodiments of the present disclosure. FIGS. 5A-5D illustrate the images taken from four different viewpoints overlaid with initial pose estimates of the object's CAD model, according to some embodiments of the present disclosure. While FIGS. 4A-4D and 5A-5D show four different viewpoints, embodiments of the present disclosure are not limited thereto and any suitable number N (e.g., two or more) of cameras 206 may capture images from the objects 300 from N different viewpoints.

The augmentation data generator 220 may receive images provided by first to fourth cameras 206 that capture different views of the object 300 from different view point (e.g., first to fourth viewpoints; see, e.g., FIGS. 4A-4D). In some embodiments, the augmentation data generator 220 supplies these images, the CAD model of the object 300, and the camera and rotation calibration information to the user device 230 that is equipped with a pose annotation tool (e.g., a multi-view pose annotation tool) 232.

According to some examples, the tool 232 allows a user to label the pose of an object 300 within a scene (e.g., an object in the top right corner of the cluttered bin, relative to a viewpoint (e.g., the top left viewpoint of FIG. 4A). The user may identify the particular object to be posed by, for example, drawing a bounding box encompassing the object. The pose estimator 224 then generates the pose estimate given the size and location of the box (FIGS. 5A-5D). The tool 232 allows the user to adjust the pose estimate by rotating or translating it on the x, y or z axis of either the camera or the cad model. Using the calibration information, the annotation tool 232 can project the pose estimate onto the different viewpoints. Thus, a user may adjust the pose estimate across different viewpoints until the refined pose appears to be a close match on all four viewpoints. FIGS. 6A-6B illustrate the refined poses across two viewpoints as overlays of key points 502, and FIGS. 7A-7B illustrate the refined poses across the two viewpoints as overlays of a rendered CAD model 304.

According to some embodiments, the augmentation data generator 220 receives this refined 6DoF pose with respect to a particular viewpoint and uses it as a ground truth pose (or label) for that viewpoint. Using the transformation matrixes from the camera calibrations, the augmentation data generator 220 can project the 6DoF pose onto the image captured by each camera 206 and use the projected 6DoF pose as the label or ground truth pose for the image. Further, by using the sensor calibration information, the augmentation data generator 220 can generate projections of the refined 6DoF pose onto the sensed data (e.g., heat map or depth map) and save the projections as labels for the sensed data.

In some embodiment, the controller 222 of the augmentation data generator 220 rotates the rotation device 204 to certain angles and creates various lighting conditions via the light sources 210. At each angle of rotation and lighting condition, the cameras 206 and sensors 208 capture the scene. For example, the cameras 206 may capture RGB and polarization images, the infrared sensors may detect the temperature variations across the target object 300 and generate a heat map, and the depth sensors may detect depth variations across the object 300 and generate a depth map accordingly. As the rotation transformation is known through the rotation calibration, the augmentation data generator 220 can reproject the refined pose received from the user device 230 to each of the captured images and sensed data at each rotation angle and lighting condition and use the reprojected pose data at each instance as the data label. Thus, according to some embodiments, by rotating the rotation device 204 to different angles, creating various lighting environments, and automatically capturing and labeling data at each rotation and lighting condition, the data capture stage 200 can automatically generate a large set of training data 101 from a single scene. In some examples, the rotation device 204 may be rotated without changing the lighting environment, and at each rotation a number of images and sensed data may be captured. Conversely, the rotation device 204 may be maintained in a fixed angle while the lighting condition is changed and the scene captured under different lighting conditions.

Using this method, the augmentation data generator 220 may aggregate a plethora of labeled data, which includes RGB images, polarization images, 3D point clouds, depth maps, heat maps, UV maps, and/or the like. Each piece of data may be labeled with one or more of a 6DoF pose, a bounding box, key points, 3D CAD model.

As shown in FIGS. 4A-4D, the target object 300 may be part of a random arrangement 302 of the same objects (e.g., part of a heavily cluttered bin of manufactured parts). By repeating the process described above for the different objects 300 within the arrangement 302 (e.g., or the different test tubes within the cluttered bin of FIGS. 4A-4D), the augmentation data generator 220 can create training data for the different objects. The objects 300 may be substantially identical but have random poses or be partially occluded by other objects. As some of these poses and occlusions may be difficult to replicate with only a single target object in a scene, the data capture stage 200 allows training data to be generated for an object in a variety of poses and conditions that may be similar to those encountered in real world environments (e.g., a manufacturing environment). This training data may then be used to train a multitude of computer vision tasks.

FIGS. 8A-8B is a flow diagram illustrating a process of capturing and augmenting training data, according to some example embodiments of the present disclosure.

Referring to FIG. 8A, in some embodiments, the augmentation data generator 220 (e.g., the stage controller 222) initiates capturing of a first image of the target object 300 by a first camera (e.g., an RGB camera) 206 of the data capture stage 200, and initiates capture of a second image of the target object 300 by a second camera 206 (e.g., an RGB camera or a polarization camera 10) of the data capture stage 200 (S802). In some examples, the first image is an RGB image and the second image is an RGB image or a polarization image. In some examples, the augmentation data generator 220 initiates a first lighting condition via a plurality of light sources 210 prior to initiating the capture of the first image, and changes the lighting condition by initiating a second lighting condition prior to initiating the capture of the second image.

In some examples, the augmentation data generator 220 then provides the first and second images, a CAD model of the target object 300, and calibration information of the first and second cameras 206 to the user device 230.

The augmentation data generator 220 receives a first pose (e.g., a refined pose from the user device 230) of the target object 300 corresponding to a first viewpoint of the first image (S804). The augmentation data generator 220 then projects the first pose onto the second image to generate a second pose of the target object corresponding to a second viewpoint of the second image (S806), which is different from the first viewpoint. The first and second poses may be six-degree-of-freedom (6DoF) poses of the target object 300.

The augmentation data generator 220 generates a first label (e.g., a first ground truth pose) associated with the first image, which includes the first pose (S808), and generates a second label (e.g., a second ground truth pose) associated with the second image, which includes the second pose (S810).

The augmentation data generator 220 then generates the training data based on the first image and its associated first label and based on the second image and the associated second label (S812). The second label may further include at least one of a CAD model of the target object 300 posed according to the second pose and key points corresponding to the CAD model.

Referring to FIG. 8B, in some embodiments, the augmentation data generator 220 (e.g., the stage controller 222) initiates sensing of mapping data of the target object 300 by a sensor (e.g., an infrared sensor, a depth sensor, or a UV sensor) 208 of the data capture stage 200 (S814). In some examples, the mapping data may be a heat map, a depth map, or an ultraviolet map corresponding to the target object.

The augmentation data generator 220 projects the first pose onto the mapping data to generate a third pose of the target object corresponding to a third viewpoint of the mapping data (S816). The augmentation data generator 220 then generates a third label associated with the mapping data, the third label including the third pose (S818). The augmentation data generator 220 augments the training data based on the mapping data and the third label (S820).

The operations performed by the constituent components of the augmentation data generator and the user device of the present disclosure may be performed by a “processing circuit” or “processor” that may include any combination of hardware, firmware, and software, employed to process data or digital signals. Processing circuit hardware may include, for example, application specific integrated circuits (ASICs), general purpose or special purpose central processing units (CPUs), digital signal processors (DSPs), graphics processing units (GPUs), and programmable logic devices such as field programmable gate arrays (FPGAs). In a processing circuit, as used herein, each function is performed either by hardware configured, i.e., hard-wired, to perform that function, or by more general-purpose hardware, such as a CPU, configured to execute instructions stored in a non-transitory storage medium. A processing circuit may be fabricated on a single printed wiring board (PWB) or distributed over several interconnected PWBs. A processing circuit may contain other processing circuits; for example, a processing circuit may include two processing circuits, an FPGA and a CPU, interconnected on a PWB.

It will be understood that, although the terms “first”, “second”, “third”, etc., may be used herein to describe various elements, components, regions, layers, and/or sections, these elements, components, regions, layers, and/or sections should not be limited by these terms. These terms are used to distinguish one element, component, region, layer, or section from another element, component, region, layer, or section. Thus, a first element, component, region, layer, or section discussed below could be termed a second element, component, region, layer, or section, without departing from the scope of the inventive concept.

The terminology used herein is for the purpose of describing particular embodiments and is not intended to be limiting of the inventive concept. As used herein, the singular forms “a” and “an” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “include”, “including”, “comprises”, and/or “comprising”, when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof. As used herein, the term “and/or” includes any and all combinations of one or more of the associated listed items. Further, the use of “may” when describing embodiments of the inventive concept refers to “one or more embodiments of the inventive concept”. Also, the term “exemplary” is intended to refer to an example or illustration.

For the purposes of this disclosure, “at least one of X, Y, and Z” and “at least one selected from the group consisting of X, Y, and Z” may be construed as X only, Y only, Z only, or any combination of two or more of X, Y, and Z, such as, for instance, XYZ, XYY, YZ, and ZZ.

It will be understood that when an element or layer is referred to as being “on”, “connected to”, “coupled to”, or “adjacent” another element or layer, it can be directly on, connected to, coupled to, or adjacent the other element or layer, or one or more intervening elements or layers may be present. When an element or layer is referred to as being “directly on,” “directly connected to”, “directly coupled to”, or “immediately adjacent” another element or layer, there are no intervening elements or layers present.

As used herein, the term “substantially,” “about,” and similar terms are used as terms of approximation and not as terms of degree, and are intended to account for the inherent variations in measured or calculated values that would be recognized by those of ordinary skill in the art.

As used herein, the terms “use”, “using”, and “used” may be considered synonymous with the terms “utilize”, “utilizing”, and “utilized”, respectively.

Further, the use of “may” when describing embodiments of the inventive concept refers to “one or more embodiments of the inventive concept.” Also, the term “exemplary” is intended to refer to an example or illustration.

While the present invention has been described in connection with certain exemplary embodiments, it is to be understood that the invention is not limited to the disclosed embodiments, but, on the contrary, is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims, and equivalents thereof.

Claims

1. A method of capturing training data in a data capture stage, the method comprising: initiating capture of a first image of a target object by a first camera of the data capture stage;initiating capture of a second image of the target object by a second camera of the data capture stage;receiving a first pose of the target object corresponding to a first viewpoint of the first image;projecting, using calibration data between the first camera and the second camera, the first pose into a coordinate system of the second camera to generate a second pose of the target object corresponding to a second viewpoint of the second image;generating a second label associated with the second image, the second label comprising the second pose; andgenerating the training data based on the second image and the second label.

2. The method of claim 1, wherein the first viewpoint is different from the second viewpoint, and wherein the first image is a red-green-blue (RGB) image and the second image is an RGB image or a polarization image.

3. The method of claim 1, wherein the first and second poses comprise six-degree-of-freedom (6DoF) poses of the target object.

4. The method of claim 1, wherein the second label further comprises at least one of a computer-aided design (CAD) model of the target object posed according to the second pose and key points corresponding to the CAD model.

5. The method of claim 1, further comprising: generating a first label associated with the first image, the first label comprising the first pose,wherein the generating the training data is further based on the first image and the first label.

6. The method of claim 1, further comprising: initiating rotation of a rotation device of the data capture stage on which the target object is placed;initiating capture of a third image of the target object, the third image corresponding to a viewpoint different from the first and second viewpoints;projecting, using calibration data between the first camera and a third camera, the first pose into the coordinate system of the third camera to generate a third pose of the target object corresponding to a third viewpoint of the third image;generating a third label associated with the third image, the third label comprising the third pose; andgenerating the training data further based on the third image and the third label.

7. The method of claim 1, further comprising: initiating a first lighting condition via a plurality of light sources of the data capture stage, wherein initiating the capture of the first image is in response to the initiating the first lighting condition; andinitiating a second lighting condition via the plurality of light sources of the data capture stage, wherein initiating the capture of the second image is in response to the initiating the second lighting condition.

8. The method of claim 1, further comprising: initiating sensing of mapping data of the target object by a sensor of the data capture stage,wherein the mapping data comprises one of a heat map corresponding to the target object, a depth map corresponding to the target object, and an ultraviolet map corresponding to the target object.

9. The method of claim 8, further comprising: projecting the first pose onto the mapping data to generate a third pose of the target object corresponding to a third viewpoint of the mapping data;generating a third label associated with the mapping data, the third label comprising the third pose; andgenerating the training data further based on the mapping data and the third label.

10. The method of claim 1, further comprising: providing the first and second images, a computer-aided design (CAD) model of the target object, and calibration information of the first and second cameras to a user device,wherein receiving the first pose of the target object comprises receiving the first pose from the user device.

11. A system comprising: one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers to cause the one or more computers to perform operations comprising:initiating capture of a first image of a target object by a first camera of a data capture stage;initiating capture of a second image of the target object by a second camera of the data capture stage;receiving a first pose of the target object corresponding to a first viewpoint of the first image;projecting, using calibration data between the first camera and the second camera, the first pose into a coordinate system of the second camera to generate a second pose of the target object corresponding to a second viewpoint of the second image;generating a second label associated with the second image, the second label comprising the second pose; andgenerating training data based on the second image and the second label.

12. The system of claim 11, wherein the first viewpoint is different from the second viewpoint, and wherein the first image is a red-green-blue (RGB) image and the second image is an RGB image or a polarization image.

13. The system of claim 11, wherein the first and second poses comprise six-degree-of-freedom (6DoF) poses of the target object.

14. The system of claim 11, wherein the second label further comprises at least one of a computer-aided design (CAD) model of the target object posed according to the second pose and key points corresponding to the CAD model.

15. The system of claim 11, wherein the operations further comprise: generating a first label associated with the first image, the first label comprising the first pose,wherein the generating the training data is further based on the first image and the first label.

16. The system of claim 11, wherein the operations further comprise: initiating rotation of a rotation device of the data capture stage on which the target object is placed;initiating capture of a third image of the target object, the third image corresponding to a viewpoint different from the first and second viewpoints;projecting the first pose onto the third image to generate a third pose of the target object corresponding to a third viewpoint of the third image;generating a third label associated with the third image, the third label comprising the third pose; andgenerating the training data further based on the third image and the third label.

17. The system of claim 11, wherein the operations further comprise: initiating a first lighting condition via a plurality of light sources of the data capture stage, wherein initiating the capture of the first image is in response to the initiating the first lighting condition; andinitiating a second lighting condition via the plurality of light sources of the data capture stage, wherein initiating the capture of the second image is in response to the initiating the second lighting condition.

18. The system of claim 11, wherein the operations further comprise: initiating sensing of mapping data of the target object by a sensor of the data capture stage,wherein the mapping data comprises one of a heat map corresponding to the target object, a depth map corresponding to the target object, and an ultraviolet map corresponding to the target object.

19. The system of claim 18, wherein the operations further comprise: projecting the first pose onto the mapping data to generate a third pose of the target object corresponding to a third viewpoint of the mapping data;generating a third label associated with the mapping data, the third label comprising the third pose; andgenerating the training data further based on the mapping data and the third label.

20. The system of claim 11, wherein the operations further comprise: providing the first and second images, a computer-aided design (CAD) model of the target object, and calibration information of the first and second cameras to a user device,wherein receiving the first pose of the target object comprises receiving the first pose from the user device.

21. One or more non-transitory computer storage media encoded with computer program instructions that when executed by one or more computers cause the one or more computers to perform operations comprising: initiating capture of a first image of a target object by a first camera of the data capture stage;initiating capture of a second image of the target object by a second camera of the data capture stage;receiving a first pose of the target object corresponding to a first viewpoint of the first image;projecting, using calibration data between the first camera and the second camera, the first pose into a coordinate system of the second camera to generate a second pose of the target object corresponding to a second viewpoint of the second image;generating a second label associated with the second image, the second label comprising the second pose; andgenerating the training data based on the second image and the second label.