Determining object information that does not explicitly appear in a media unit signature

Abstract

There may be provided a method for determining object information related to an object that appears in a media element, the method may include (a) receiving or generating a signature of the media element, wherein the signature may include identifiers, wherein the identifiers may include an object identifier related to the object, wherein the object identifier points to locations of output elements, within an output of a convolutional neural network, of elements of significance in relation to the object; (b) finding a group of contributing portions of the media element that had a significant contribution to the generation, by the convolutional neural network, of the output elements of significance in relation to the object; and (c) determining object information related to the object based on a distribution of the group of the contributing portions of the media element.

Description

BACKGROUND

In object detection there is a tension between the task of object classification (“is it a pedestrian?”) and localization (“where is the pedestrian? what is its size/scale?”).

While classification task benefits from scale-invariant representations (pedestrians of all sizes have the same representation therefore it is easy to identify them), the localization task needs to preserve the scale of objects in the image

There is a growing need to provide efficient manners to provide object detection solutions that may solve this tension.

SUMMARY

A method for determining object information related to an object that appears in a media element, the method may include receiving or generating a signature of the media element, wherein the signature may include identifiers, wherein the identifiers may include an object identifier related to the object, wherein the object identifier points to locations of output elements, within an output of a convolutional neural network, of elements of significance in relation to the object; finding a group of contributing portions of the media element that had a significant contribution to the generation, by the convolutional neural network, of the output elements of significance in relation to the object; and determining object information related to the object based on a distribution of the group of the contributing portions of the media element.

The signature may include a group of contributing portions of the media element that had a significant contribution to the generation, by the convolutional neural network, of the output elements of significance in relation to the object.

The method may include generating the signature in an iterative manner, wherein each iteration may include maintaining a mapping between (a) elements of an output of the iteration, and (b) contributing portions of the media element that had a significant contribution to a generation of the elements of the output of the iteration.

The method may include maintaining mapping related only to output elements of significance of an output of the convolutional neural network.

The object information may be scale information.

The object information may be orientation information.

The finding of the group of contributing portions of the media element may include multiple search iterations; wherein each search iteration may be applied on significance contributing output elements of a current layer of the convolutional neural network; wherein each search iteration may include searching, for each one of the significance contributing output elements of the current layer for a group of significant contribution output elements of a previous layer of the convolutional neural network.

Each search iteration may include applying, at each layer of the convolutional neural network, a search function that inverses an operation of the layer of the convolutional neural network.

The operation of the current layer of the convolutional neural network may be a maxpool operation for selecting a local maximum out of a subgrid of output elements of the previous layer; and wherein the search function searches for the output element of the previous layer that had a value of the subgrid.

The operation of the current layer of the convolutional neural network may be applying a convolution kernel on a subgrid of output elements of the previous layer; and wherein the search function searches a product out of a multiplication of each output element of the sub-grid by a weight of the convolution kernel.

The media unit may be an image.

The output of the convolutional neural network may be a tensor that may include a vector of values per each segment of the image; wherein the object appears in a certain segment and wherein the object identifier points to locations of significant values, within a certain vector associated with the certain segment; and wherein the portions of the media element may be image pixels.

There may be provided a non-transitory computer readable medium for determining object information related to an object that appears in a media element, the non-transitory computer readable medium may store instructions receiving or generating a signature of the media element, wherein the signature may include identifiers, wherein the identifiers may include an object identifier related to the object, wherein the object identifier points to locations of output elements, within an output of a convolutional neural network, of elements of significance in relation to the object; finding a group of contributing portions of the media element that had a significant contribution to the generation, by the convolutional neural network, of the output elements of significance in relation to the object; and determining object information related to the object based on a distribution of the group of the contributing portions of the media element.

The signature may include a group of contributing portions of the media element that had a significant contribution to the generation, by the convolutional neural network, of the output elements of significance in relation to the object.

There may be provided a non-transitory computer readable medium that may store instructions for generating the signature in an iterative manner, wherein each iteration may include maintaining a mapping between (a) elements of an output of the iteration, and (b) contributing portions of the media element that had a significant contribution to a generation of the elements of the output of the iteration.

The non-transitory computer readable medium that may store instructions for maintaining mapping related only to output elements of significance of an output of the convolutional neural network.

The object information may be scale information.

The object information may be orientation information.

The finding of the group of contributing portions of the media element may include multiple search iterations; wherein each search iteration may be applied on significance contributing output elements of a current layer of the convolutional neural network; wherein each search iteration may include searching, for each one of the significance contributing output elements of the current layer for a group of significant contribution output elements of a previous layer of the convolutional neural network.

Each search iteration may include applying, at each layer of the convolutional neural network, a search function that inverses an operation of the layer of the convolutional neural network.

The operation of the current layer of the convolutional neural network may be a maxpool operation for selecting a local maximum out of a subgrid of output elements of the previous layer; and wherein the search function searches for the output element of the previous layer that had a value of the subgrid.

The operation of the current layer of the convolutional neural network may be applying a convolution kernel on a subgrid of output elements of the previous layer; and wherein the search function searches a product out of a multiplication of each output element of the sub-grid by a weight of the convolution kernel.

The media unit may be an image.

The output of the convolutional neural network may be a tensor that may include a vector of values per each segment of the image; wherein the object appears in a certain segment and wherein the object identifier points to locations of significant values, within a certain vector associated with the certain segment; and wherein the portions of the media element may be image pixels.

There may be provided a processing circuit for executing at least one of the methods disclosed in the specification and/or the claims and/or the drawings. The processing circuitry may be implemented as a central processing unit (CPU), and/or one or more other integrated circuits such as application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), full-custom integrated circuits, etc., or a combination of such integrated circuits.

The processing circuit may be configured to determining object information related to an object that appears in a media element, by (a) receiving or generating a signature of the media element, wherein the signature may include identifiers, wherein the identifiers may include an object identifier related to the object, wherein the object identifier points to locations of output elements, within an output of a convolutional neural network, of elements of significance in relation to the object; (b) finding a group of contributing portions of the media element that had a significant contribution to the generation, by the convolutional neural network, of the output elements of significance in relation to the object; and (c) determining object information related to the object based on a distribution of the group of the contributing portions of the media element.

BRIEF DESCRIPTION OF THE DRAWINGS

The embodiments of the disclosure will be understood and appreciated more fully from the following detailed description, taken in conjunction with the drawings in which:

FIG. 1 illustrates an example of a generation of a signature of an image by a convolutional neural network (CNN);

FIG. 2 illustrates an example of segments of the image, an output of the CNN and the signature of the image;

FIG. 3 illustrates a process for obtaining object information from the signature of the image;

FIG. 4 illustrates a process for obtaining object information from the signature of the image;

FIG. 5 is an example of a method;

FIG. 6 illustrates an example of an image with two objects of different scale; and

FIG. 7 illustrates an example of an image with two objects of different scale; and

FIG. 8 illustrates an example of an image with objects of different scales and orientations.

DESCRIPTION OF EXAMPLE EMBODIMENTS

In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be understood by those skilled in the art that the present invention may be practiced without these specific details. In other instances, well-known methods, procedures, and components have not been described in detail so as not to obscure the present invention.

The subject matter regarded as the invention is particularly pointed out and distinctly claimed in the concluding portion of the specification. The invention, however, both as to organization and method of operation, together with objects, features, and advantages thereof, may best be understood by reference to the following detailed description when read with the accompanying drawings.

It will be appreciated that for simplicity and clarity of illustration, elements shown in the figures have not necessarily been drawn to scale. For example, the dimensions of some of the elements may be exaggerated relative to other elements for clarity. Further, where considered appropriate, reference numerals may be repeated among the figures to indicate corresponding or analogous elements.

Because the illustrated embodiments of the present invention may for the part, be implemented using electronic components and circuits known to those skilled in the art, details will not be explained in any greater extent than that considered necessary as illustrated above, for the understanding and appreciation of the underlying concepts of the present invention and in order not to obfuscate or distract from the teachings of the present invention.

Any reference in the specification to a method should be applied mutatis mutandis to a device or system capable of executing the method and/or to a non-transitory computer readable medium that stores instructions for executing the method.

Any reference in the specification to a system or device should be applied mutatis mutandis to a method that may be executed by the system, and/or may be applied mutatis mutandis to non-transitory computer readable medium that stores instructions executable by the system.

Any reference in the specification to a non-transitory computer readable medium should be applied mutatis mutandis to a device or system capable of executing instructions stored in the non-transitory computer readable medium and/or may be applied mutatis mutandis to a method for executing the instructions.

Any combination of any module or unit listed in any of the figures, any part of the specification and/or any claims may be provided.

There may be provided a system, non-transitory computer readable medium and method for extracting object information such as a scale information, orientation information or other information from a signature of an image that does not explicitly include that image information.

Thus—the signature itself may be compact and robust to changes in scale, to changes in orientation or to other changes—but the scale information or orientation information may be extracted.

Using such a signature increases the accuracy of object detection, required less storage and processing resources. For example—less signatures may be stored (as there is no need in a signature per scale, per orientation, and the like), in order to match signatures fewer processing resources should be allocated.

The object information may be at least one out of (a) any statistics (for example—aggregated statistics) regarding the contributing portions of the media element (the statistics may be applied on a pixel basis or on a coarser basis), (b) precise location of the contributing portions, for example—in a case of an image size of a contributing area aka “scale” and the orientation of the area, (c) in case of an image-predominant color of the pixels in the contributing portion, (d) apply additional filters/layers only to specified contributing portions. For example—the method may find a particular identifier, find a contributing portion associated with the identifier and then apply a smaller auxiliary network only to this contributing portion. This will be much more computationally efficient than applying this auxiliary network to the whole media unit.

In the following text it is assumed that the media unit is an image—although the media unit may be an audio unit, a multimedia unit, any representation of a measurement of a natural signals, any other measurements, and the like.

The following text may refer to a significant value or to a significant contribution. The significant value may be the most significant value and the significant contribution may be the highest contribution. A significant value is a value that fulfills a significant criterion. The significant value may be a value that exceeds a threshold, a value that is within a predefined value range, belong to a group of highest values, have a highest absolute value, be below a certain threshold, be of a smallest value, belong to a group of lowest values, and the like. The significance criterion may be determined by the function applied by one or more layers of the CNN, may be determined based on a response of one or more layers of the CNN to objects that were marked or determined as significant, and the like.

FIG. 1 illustrates an example of a signature generation process.

An image 6001 is fed to a convolutional network layer (CNN) 6010 that has multiple (N) layers 6010(1)-6010(N), N being a positive integer that may exceed three.

Each layer applies a function and the CNN applies N functions 6015(1)-6015(N). The output of the N'th layer 6010(N) is denoted CNN output 6018. The CNN output is further processed (by step 6020 of a completion of signature generation) to provide an image signature 6028.

CNN output 6018 may include multiple values per each segment of the media unit. The values may represent attributes, properties, and the like.

Step 6020 of completion of signature generation may include selecting, out of values of the CNN output the significant values per media unit segment.

FIG. 2 illustrates an example of CNN output 6018 and of image signature 6028.

The image 6001 is virtually segments to segments 6000(i,k). The segments may be of the same shape and size but this is not necessarily so. One or more segments may at least partially overlap and/or one or more segments may not overlap.

CNN output 6018 is a tensor that includes a vector of values per each segment of the media unit. An object appears in a certain segment. An object identifier (of the signature) points to locations of significant values, within a certain vector associated with the certain segment.

For example—a top left segment (6001(1,1)) of the image may be represented in the CNN output 6018 by a vector V(1,1) 6018(1,1) that has multiple values. The number of values per vector may exceed 100, 200, 500, 1000, and the like.

The significant values (for example—more than 10, 20, 30, 40 values, and/or more than 0.1%, 0.2%. 0.5%, 1%, 5% of all values of the vector and the like) may be selected. The significant values may have the values—but may be selected in any other manner.

FIG. 2 illustrates a set of significant responses 6016(1,1) of vector V(1,1) 6018(1,1). The set includes five significant values (such as first significant value SV1(1,1) 6014(1,1,1), second significant value SV2(1,1), third significant value SV3(1,1), fourth significant value SV4(1,1), and fifth significant value SV5(1,1) 6014(1,1,5).

The image signature 6028 includes five indexes for the retrieval of the five significant values—first till fifth identifiers ID1-ID5 are indexes for retrieving the first till fifth significant values.

FIG. 3 illustrates an example of a signature generation process that include maintaining information that links pixels of the image 6001 to the output of each one of the layers of the CNN, and links pixels of the image to the values of the CNN output that are identified by indexes of the signature.

Once the values of the CNN output are selected and the selected values are identified by identifiers in the signature—information related to pixels that are related to non-elected values may be deleted—to save storage space.

Thus—the outcome of the N layers of the CNN network (6010(1)-6010(N)) also includes the pixels that are linked to the output of each of the layer —6018(1)-6018(N).

FIG. 4 illustrates a process for obtaining object information from the signature of the image.

The process includes tracing back, for each layer (of the second till N'th layer) of the CNN the outputs of a previous layer that contributed to the significant values of the current layer. For the first layer—tracing back the pixels that contributed to the significant values of the output of the first CNN layer.

Thus N traceback (TB) functions are applied—from first TB function till N'th TB function 6017(1)-6017(N).

Thus, for a first identifier ID1 in the signature—(identifier first significant value SV16014(1,1)), the N'th TB function 6017(N) finds the set of contributing (N−1)'th layer outputs 6013(N−1). The first TB function 6017(1) finds the pixels (set of pixels) related to ID16001(9)—by searching pixels that contributed to significant first layer output values 6013(1).

FIG. 5 illustrates an example of a first image 6001 that includes a standing first person 6051 of a first scale and a standing person 6052 of a second scale that differs from the first scale.

FIG. 5 also illustrates an example of a second image 6002 that includes a laying first person 6051 of a first scale and a standing person 6052 of a second scale that differs from the first scale.

The first and second images are segmented to segments and each segment may be represented by a vector, out of which significant values will be selected and identified by identifiers (indexes) that belong to the first image signature and the second image signature respectively.

The first person 6051 will be identified by first index ID116061 in the signatures of the first and second images and the second person 6052 will be identified by second index ID126062 in the signatures of the first and second images. These signature do not include (explicitly) the scale and orientation of the first and second persons—but the methods disclosed in the application extract the scale and orientation information.

FIG. 6 illustrates an example of a method 6100.

Method 6100 includes a sequence of steps 6102, 6104, 6106, 6108, 6110, and 6112.

Step 6102 may include selecting an index out of multiple indices of an image signature, wherein the image signature lacks explicit scale fields.

The selection can be made according to various criteria. For example—the selection may follow a match between a concept structure and a signature of an image. The concept structure may include signatures and metadata related to the signatures. A concept may represent an object.

Step 6102 may include steps 6102(1), 6102(2) and 6102(3).

Step 6102(1) may include receiving or generating a signature of an image 6102(1).

Step 6102(2) may include checking if the signature of the image matches a signature of a concept structure that includes a collection of signatures. A match occurs if at least a predefined number of identifies in the signature of the image match identifiers of a signature of the concept structure.

If a matching occurs—selecting (6102(3)) the matching signatures to undergo steps 6104-6112.

Step 6104 may include using the index to find a set of significant responses in a tensor that is outputted from a last (N'th) layer of a convolutional neural network (CNN).

Step 6106 may include, for each significant response of the set—tracing back the set of contributing output elements of the (N−1)th layer of the CNN.

Step 6108 may include performing, for each layer of the CNN, out of the (N−1) layer to the second layer of the CNN, finding, for each contributing output element of the set of contributing output elements of a current layer, the set of contributing output elements of the pervious layer.

Step 6110 may include finding, for each contributing output element of the set of contributing output elements of the first layer, the pixels of the image that are the contributing pixels to the contributing output element.

Step 6112 may include determining one or more property of the object associated with the selected index based on the distribution of the contributing pixels.

FIG. 7 illustrates method 6200 for determining object information related to an object that appears in a media element.

Method 6200 may include steps 6202, 6204 and 6206.

Step 6202 may include receiving or generating a signature of the media element, wherein the signature may include identifiers. The identifiers may include an object identifier related to the object. The object identifier points to locations of output elements, within an output of a convolutional neural network, of elements of significance in relation to the object.

The output of the convolutional neural network may be a tensor that may include a vector of values per each segment of the image; wherein the object appears in a certain segment and wherein the object identifier points to locations of significant values, within a certain vector associated with the certain segment; and wherein the portions of the media element may be image pixels. Step 6204 may applied on selected identifiers—such as identifier that match a signature of a concept data structure.

Step 6204 may include finding a group of contributing portions of the media element that had a significant contribution to the generation, by the convolutional neural network, of the output elements of significance in relation to the object.

Step 6206 may include determining object information related to the object based on a distribution of the group of the contributing portions of the media element.

The signature may include a group of contributing portions of the media element that had a significant contribution to the generation, by the convolutional neural network, of the output elements of significance in relation to the object.

Step 6202 may include generating the signature in an iterative manner, wherein each iteration may include maintaining a mapping between (a) elements of an output of the iteration, and (b) contributing portions of the media element that had a significant contribution to a generation of the elements of the output of the iteration. In this case step 6204 may included in step 6202 or may include retrieving the information generated in step 6204.

Step 6202 may include maintaining mapping related only to output elements of significance of an output of the convolutional neural network.

The object information may be scale information.

The object information may be orientation information.

Step 6204 may include multiple search iterations; wherein each search iteration may be applied on significance contributing output elements of a current layer of the convolutional neural network; wherein each search iteration may include searching, for each one of the significance contributing output elements of the current layer for a group of significant contribution output elements of a previous layer of the convolutional neural network.

Each search iteration may include applying, at each layer of the convolutional neural network, a search function that inverses an operation of the layer of the convolutional neural network.

The operation of the current layer of the convolutional neural network may be a maxpool operation for selecting a local maximum out of a subgrid of output elements of the previous layer; and wherein the search function searches for the output element of the previous layer that had a value of the subgrid.

The operation of the current layer of the convolutional neural network may be applying a convolution kernel on a subgrid of output elements of the previous layer; and wherein the search function searches a product out of a multiplication of each output element of the sub-grid by a weight of the convolution kernel.

FIG. 8 illustrates an example of an image with an object at different scale and orientation.

The object 6300 should be identified by the same identifier—but its distribution differs as a function of scale and orientation.

The distribution of the pixels related to each of the object may represent an approximation of the border of the objects and may provide information about properties such as scale and orientation. The orientation may be determined in various manners—for example find axes that cross the object and may pass through a center of gravity of the object and determine the orientation of the object based on the orientation of the axes. The orientation may be determined during a learning period, based on an equation or a lookup table.

While the foregoing written description of the invention enables one of ordinary skill to make and use what is considered presently to be the best mode thereof, those of ordinary skill will understand and appreciate the existence of variations, combinations, and equivalents of the specific embodiment, method, and examples herein. The invention should therefore not be limited by the above described embodiment, method, and examples, but by all embodiments and methods within the scope and spirit of the invention as claimed.

In the foregoing specification, the invention has been described with reference to specific examples of embodiments of the invention. It will, however, be evident that various modifications and changes may be made therein without departing from the broader spirit and scope of the invention as set forth in the appended claims.

Moreover, the terms “front,” “back,” “top,” “bottom,” “over,” “under” and the like in the description and in the claims, if any, are used for descriptive purposes and not necessarily for describing permanent relative positions. It is understood that the terms so used are interchangeable under appropriate circumstances such that the embodiments of the invention described herein are, for example, capable of operation in other orientations than those illustrated or otherwise described herein.

Furthermore, the terms “assert” or “set” and “negate” (or “deassert” or “clear”) are used herein when referring to the rendering of a signal, status bit, or similar apparatus into its logically true or logically false state, respectively. If the logically true state is a logic level one, the logically false state is a logic level zero. And if the logically true state is a logic level zero, the logically false state is a logic level one.

Those skilled in the art will recognize that the boundaries between logic blocks are merely illustrative and that alternative embodiments may merge logic blocks or circuit elements or impose an alternate decomposition of functionality upon various logic blocks or circuit elements. Thus, it is to be understood that the architectures depicted herein are merely exemplary, and that in fact many other architectures may be implemented which achieve the same functionality.

Any arrangement of components to achieve the same functionality is effectively “associated” such that the desired functionality is achieved. Hence, any two components herein combined to achieve a particular functionality may be seen as “associated with” each other such that the desired functionality is achieved, irrespective of architectures or intermedial components. Likewise, any two components so associated can also be viewed as being “operably connected,” or “operably coupled,” to each other to achieve the desired functionality.

Furthermore, those skilled in the art will recognize that boundaries between the above described operations merely illustrative. The multiple operations may be combined into a single operation, a single operation may be distributed in additional operations and operations may be executed at least partially overlapping in time. Moreover, alternative embodiments may include multiple instances of a particular operation, and the order of operations may be altered in various other embodiments.

Also for example, in one embodiment, the illustrated examples may be implemented as circuitry located on a single integrated circuit or within a same device. Alternatively, the examples may be implemented as any number of separate integrated circuits or separate devices interconnected with each other in a suitable manner.

However, other modifications, variations and alternatives are also possible. The specifications and drawings are, accordingly, to be regarded in an illustrative rather than in a restrictive sense.

In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word ‘comprising’ does not exclude the presence of other elements or steps then those listed in a claim. Furthermore, the terms “a” or “an,” as used herein, are defined as one or more than one. Also, the use of introductory phrases such as “at least one” and “one or more” in the claims should not be construed to imply that the introduction of another claim element by the indefinite articles “a” or “an” limits any particular claim containing such introduced claim element to inventions containing only one such element, even when the same claim includes the introductory phrases “one or more” or “at least one” and indefinite articles such as “a” or “an.” The same holds true for the use of definite articles. Unless stated otherwise, terms such as “first” and “second” are used to arbitrarily distinguish between the elements such terms describe. Thus, these terms are not necessarily intended to indicate temporal or other prioritization of such elements. The mere fact that certain measures are recited in mutually different claims does not indicate that a combination of these measures cannot be used to advantage.

While certain features of the invention have been illustrated and described herein, many modifications, substitutions, changes, and equivalents will now occur to those of ordinary skill in the art. It is, therefore, to be understood that the appended claims are intended to cover all such modifications and changes as fall within the true spirit of the invention.

It is appreciated that various features of the embodiments of the disclosure which are, for clarity, described in the contexts of separate embodiments may also be provided in combination in a single embodiment. Conversely, various features of the embodiments of the disclosure which are, for brevity, described in the context of a single embodiment may also be provided separately or in any suitable sub-combination.

It will be appreciated by persons skilled in the art that the embodiments of the disclosure are not limited by what has been particularly shown and described hereinabove. Rather the scope of the embodiments of the disclosure is defined by the appended claims and equivalents thereof.

Claims

1. A method for determining object information related to an object that appears in a media element, the method comprises: receiving or generating a signature of the media element, wherein the signature comprises identifiers, wherein the identifiers comprise an object identifier related to the object, wherein the object identifier points to locations of output elements, within an output of a convolutional neural network, of elements of significance in relation to the object;finding a group of contributing portions of the media element that had a significant contribution to the generation, by the convolutional neural network, of the output elements of significance in relation to the object; wherein group of contribution portions of the media elements that had the significant contribution comprises a certain number of portions of the media element that had a higher contribution than other portions of the media element;wherein the finding of the group of contributing portions of the media element comprises multiple search iterations;wherein each search iteration is applied on significance contributing output elements of a current layer of the convolutional neural network;wherein each search iteration comprises searching, for each one of the significance contributing output elements of the current layer, for a group of significant contribution output elements of a previous layer of the convolutional neural network;anddetermining object information related to the object based on a distribution of the group of the contributing portions of the media element.

2. The method according to claim 1 wherein the signature comprises a group of contributing portions of the media element that had a significant contribution to the generation, by the convolutional neural network, of the output elements of significance in relation to the object.

3. The method according to claim 1 comprising generating the signature in an iterative manner, wherein each iteration comprises maintaining a mapping between (a) elements of an output of the iteration, and (b) contributing portions of the media element that had a significant contribution to a generation of the elements of the output of the iteration.

4. The method according to claim 3 comprising maintaining mapping related only to output elements of significance of an output of the convolutional neural network.

5. The method according to claim 1 wherein the object information is scale information.

6. The method according to claim 1 wherein the object information is orientation information.

7. The method according to claim 1 wherein the certain number is a predefined number or a certain percent of all portions of the media element.

8. The method according to claim 1 wherein each search iteration comprising applying, at each layer of the convolutional neural network, a search function that inverses an operation of the layer of the convolutional neural network.

9. The method according to claim 8 wherein the operation of the current layer of the convolutional neural network is a maxpool operation for selecting a local maximum out of a subgrid of output elements of the previous layer; and wherein the search function searches for the output element of the previous layer that had a value of the subgrid.

10. The method according to claim 8 wherein the operation of the current layer of the convolutional neural network is applying a convolution kernel on a subgrid of output elements of the previous layer; and wherein the search function searches a product out of a multiplication of each output element of the sub-grid by a weight of the convolution kernel.

11. The method according to claim 1 wherein the media unit is an image.

12. The method according to claim 11 wherein the output of the convolutional neural network is a tensor that comprises a vector of values per each segment of the image; wherein the object appears in a certain segment and wherein the object identifier points to locations of significant values, within a certain vector associated with the certain segment; and wherein the portions of the media element are image pixels.

13. A non-transitory computer readable medium for determining object information related to an object that appears in a media element, the non-transitory computer readable medium stores instructions: receiving or generating a signature of the media element, wherein the signature comprises identifiers, wherein the identifiers comprise an object identifier related to the object, wherein the object identifier points to locations of output elements, within an output of a convolutional neural network, of elements of significance in relation to the object;finding a group of contributing portions of the media element that had a significant contribution to the generation, by the convolutional neural network, of the output elements of significance in relation to the object;wherein group of contribution portions of the media elements that had the significant contribution comprises a certain number of portions of the media element that had a higher contribution than other portions of the media element;wherein the finding of the group of contributing portions of the media element comprises multiple search iterations;wherein each search iteration is applied on significance contributing output elements of a current layer of the convolutional neural network;wherein each search iteration comprises searching, for each one of the significance contributing output elements of the current layer, for a group of significant contribution output elements of a previous layer of the convolutional neural network;anddetermining object information related to the object based on a distribution of the group of the contributing portions of the media element.

14. The non-transitory computer readable medium according to claim 13 wherein the signature comprises a group of contributing portions of the media element that had a significant contribution to the generation, by the convolutional neural network, of the output elements of significance in relation to the object.

15. The non-transitory computer readable medium according to claim 13 that stores instructions for generating the signature in an iterative manner, wherein each iteration comprises maintaining a mapping between (a) elements of an output of the iteration, and (b) contributing portions of the media element that had a significant contribution to a generation of the elements of the output of the iteration.

16. The non-transitory computer readable medium according to claim 15 that stores instructions for maintaining mapping related only to output elements of significance of an output of the convolutional neural network.

17. The non-transitory computer readable medium according to claim 13 wherein the object information is scale information.

18. The non-transitory computer readable medium according to claim 13 wherein the object information is orientation information.

19. The non-transitory computer readable medium according to claim 13 wherein the wherein the certain number is a predefined number or a certain percent of all portions of the media element.

20. The non-transitory computer readable medium according to claim 13 wherein each search iteration comprising applying, at each layer of the convolutional neural network, a search function that inverses an operation of the layer of the convolutional neural network.

21. The non-transitory computer readable medium according to claim 20 wherein the operation of the current layer of the convolutional neural network is a maxpool operation for selecting a local maximum out of a subgrid of output elements of the previous layer; and wherein the search function searches for the output element of the previous layer that had a value of the subgrid.

22. The non-transitory computer readable medium according to claim 20 wherein the operation of the current layer of the convolutional neural network is applying a convolution kernel on a subgrid of output elements of the previous layer; and wherein the search function searches a product out of a multiplication of each output element of the sub-grid by a weight of the convolution kernel.

23. The non-transitory computer readable medium according to claim 13 wherein the media unit is an image.

24. The non-transitory computer readable medium according to claim 23 wherein the output of the convolutional neural network is a tensor that comprises a vector of values per each segment of the image; wherein the object appears in a certain segment and wherein the object identifier points to locations of significant values, within a certain vector associated with the certain segment; and wherein the portions of the media element are image pixels.