METHOD AND SYSTEM FOR POSE ESTIMATION

Abstract

The disclosures relate to a method and a system for pose estimation. The method comprises: extracting a plurality of sets of part-feature maps from an image, each set of the extracted part-feature maps encoding the messages for a particular body part and forming a node of a part-feature network; passing a message of each set of the extracted part-feature maps through the part-feature network to update the extracted part-feature maps, resulting in each set of the extracted part-feature maps incorporating the message of upstream nodes; estimating, based on the updated part-feature maps, the body part within the image.

Description

TECHNICAL FIELD

The disclosures relate to a method and a system for pose estimation.

BACKGROUND

Human pose estimation is to estimate locations of body parts from images, which could be applied to a variety of vision tasks, such as action recognition, tracking, and human-computer interaction. Despite the long history of efforts, pose estimation is still a challenging and unsolved problem. The large variations in limb orientation, clothing, viewpoints, background clutters, truncation, and occlusion make pose estimation even more difficult.

It is known that pose estimation has been achieved by refining appearance score maps generated from input images. Regarding the score maps, the information on visual patterns at a location is summarized into a single probability value, indicating the likelihood of the existence of the corresponding body part. For example, If a location of an elbow has a large response on the score map, it can be concluded that this location may belong to elbow. However, the in-plane and out-plane rotation of the elbow, the orientations of the upper arm and the lower arm associated with it, the occlusion status, etc., are inconclusive. Such detailed information is valuable for estimating the locations of other body parts, but is lost from the score map, which makes structural learning among body parts less effective.

SUMMARY

The following presents a simplified summary of the disclosure in order to provide a basic understanding of some aspects of the disclosure. This summary is not an extensive overview of the disclosure. This summary neither identifies key or critical elements of the disclosure nor delineates any scope of particular embodiments of the disclosure, or any scope of the claims. Its sole purpose is to present some concepts of the disclosure in a simplified form as a prelude to the more detailed description that is presented later.

It is observed that the inter-part information is well preserved at the feature level, where hierarchical feature representations are learned with Convolutional Neural Networks. In order to address, at least partially, one of the above issues, a method is proposed for pose estimation, in one aspect of the present application. The method comprises: extracting a plurality of sets of part-feature maps from an image, each set of the extracted part-feature maps representing a body part and forming a node of a part-feature network; passing a message of each set of the extracted part-feature maps through the part-feature network to update the extracted part-feature maps, resulting in each set of the extracted part-feature maps incorporating the message of upstream nodes; estimating, based on the updated part-feature maps, the body part within the image. The pose estimation accuracy will be enhanced since rich information contained in part-feature maps between body parts can be mostly preserved as a result of message passing process in the feature level.

In one embodiment of the present application, the message is passed in opposite directions and each pairs of the updated part-feature maps performed in different directions are combined into a score map. The generated score map is used to estimate poses within the image with improved accuracy.

In one embodiment of the present application, the part-feature maps are extracted via a CNN, and preferably, a VGG net.

In one embodiment of the present application, only 3 pooling layers of the VGG net are enabled in order to preserve a higher resolution.

In one embodiment of the present application, the message is passed through a convolution operation with a geometrical transformation kernel.

In another aspect, a system for pose estimation is provided. The system comprises: a memory that stores executable components; and a processor electrically coupled to the memory to execute the executable components for: extracting a plurality of sets of part-feature maps from an image, each set of the extracted part-feature maps representing a body part and forming a node of a part-feature network; passing, node by node, a message of each set of the extracted part-feature maps through the part-feature network to update the extracted part-feature maps, resulting in each set of the extracted part-feature maps incorporating the message of previously passed nodes; estimating, based on the refined part-feature maps, the body part within the image.

BRIEF DESCRIPTION OF THE DRAWING

Exemplary non-limiting embodiments of the present application are described below with reference to the attached drawings. The drawings are illustrative and generally not to an exact scale. The same or similar elements on different figures are referenced with the same reference numbers.

FIG. 1 illustrates an exemplary system for estimating poses from an input image according to one embodiment of the present application;

FIG. 2 is a flow chart illustrating a process for estimating poses from an input image according to one embodiment of the present application;

FIG. 3 illustrates another exemplary system for estimating poses from an input image according to one embodiment of the present application;

FIG. 4 illustrates a geometrical transformation process according to one embodiment of the present application;

FIG. 5 illustrates a feature map updating operation according to one embodiment of the present application; and

FIG. 6 illustrates a process of a bi-directional message passing according to one embodiment of the present application.

DETAILED DESCRIPTION

Reference will now be made in detail to some specific embodiments of the present application contemplated by the inventors for carrying out the present application. Examples of these specific embodiments are illustrated in the accompanying drawings. While the present application is described in conjunction with these specific embodiments, it will be appreciated by one skilled in the art that it is not intended to limit the present application to the described embodiments. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. The present application may be practiced without some or all of these specific details. In other instances, well-known process operations have not been described in detail in order not to unnecessarily obscure the present application.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used herein, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “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.

An exemplary system 1000 for estimating poses from an input image will now be described with reference to FIG. 1. A Convolutional Neural Network (CNN) 1200 is utilized to generate semantically meaningful representations from an input image 1100. It is appreciated that the CNN 1200 may employ multiple layers 1210 to learn hierarchical feature representations of input images. Features in lower layers capture low-level information, while those in higher layers can represent more abstract concepts, such as attributes and object categories. In this embodiment, the CNN 1200 employs, for example, 7 layers. Widely used CNNs, such as AlexNet, Clarifai, Overfeat, GoogleNet, and Visual Geometry Group (VGG), employ fully connected (fc) layers following convolutional layers to obtain global feature vectors. In Fully Convolutional Nets (FCNs), a 1×1 convolution framework is used to replace the fc layers. In this embodiment, a fully convolutional VGG net is preferably chosen as a base model. In a preferred embodiment, only 3 pooling layers out of 5 are enabled in order to preserve a higher resolution. A feature vector obtained at location (x, y) in layer fcn6 is denoted by h_fcn6(x,y). The feature vector for body part k at location (x, y) extracted from layer fcn6 is computed as

h_fcn7^k(x,y)=f(h_fcn6(x,y)⊗w_fcn7^k)  (1)

wherein ⊗denotes a convolution operation, f denotes a nonlinear function, and w_fcn7^k denotes a filter bank for part k. It should be noted that, h_fcn7^k contains a set of part-feature maps extracted from different channels. The part-feature maps of body parts contain rich information and detailed descriptions of human poses and appearance.

Since spatial distributions and co-occurrence of part-feature maps obtained at different parts are highly correlated, passing the rich information contained in part-feature maps between parts can effectively improve features learned at each part. In the prior art, the passing process is implemented in the score map level, which results in the loss of important inter-part information. Surprisingly, when a message passes through at the feature level, the rich information contained in part-feature maps between parts is largely preserved.

In the present application, the geometric constraints among body parts could be consolidated by shifting part-feature maps of one body part towards neighboring parts. The geometrical transformation kernels model the relationship between every pair of part-feature maps from neighboring parts. To optimize features obtained at a part, it is expected to receive information from all other parts with a fully connected graph. However, in order to directly model the relationship between part-feature maps of parts in distance, large transformation kernels, which are difficult to be trained, have to be introduced. Second, the relationships between some parts (such as head and foot) are unstable. It is advantageous to pass message between them through intermediate parts on a designed graph, since the relative spatial distribution between the two adjacent parts is stable and the corresponding kernel is easy to be trained. The adjacent parts on the graph are close in distance and have relatively stable relationship in the graph. The extracted sets of part-feature maps constitute a part-feature network processed by a structured feature learning layer 1220, wherein each set of part-feature maps occupies a node 1221 in the part-feature network. In an exemplary implement, a message of each set of the extracted part-feature maps is passed through the part-feature network along a unitary direction. The passing operation will be illustrated in detail with reference to FIG. 6. After the message passing, each set of the part-feature maps are updated through message passing in the part-feature network. As a result, each set of the part-feature maps incorporates the message of previously passed nodes, except for the part-feature map occupying the leaf node. In a preferred embodiment, the updated part-feature maps are transformed (linearly combined, for example) into score maps from which the part positions can be estimated. With the updated part-feature maps, the pose estimation can achieve a higher accuracy.

The flow chart illustrating a process for estimating poses from an input image is schematically shown in FIG. 2. A plurality of sets of part-feature maps are extracted, through a CNN, from an input image at step S2020. The extracted part-feature maps are sorted into a plurality of sets representing various parts of a human body, respectively. In an exemplary embodiment, the extracted sets of part-feature maps constitute a body-like network. A message of each set of the extracted part-feature maps is passed through the network at step 2040. In a preferred embodiment, the updated part-feature maps are linearly combined into score maps at step S2060. Based on the score maps, the poses within the image are detected out at step S2080.

Referring to FIG. 3, an alternative embodiment of pose estimation system 3000 is provided. The system 3000 is similar to the above-mentioned system 1000 except for the structured feature learning layer 3220. In this embodiment, the message is passed bi-directionally through the part-feature network, which enables the extracted part-feature maps incorporating the message of adjacent nodes in two directions. The updated pairs of part-feature maps in opposite directions are combined into score maps which further improve the pose estimation accuracy.

FIG. 4 and FIG. 5 illustrate a detailed geometrical transformation operation that can be utilized in message passing. As shown in FIG. 4, a feature map 4100 under Gaussian distribution is provided. Three different geometrical transformation kernels 4200 are convoluted with the same feature map 4100, respectively, resulting differently transformed feature maps 4300. As can be seen from FIG. 4, the transformed feature maps have been shifted towards top-left, top-right, and bottom-left, respectively. In pose estimation, the geometric constraints among body parts can be consolidated by shifting the feature map of one body part towards adjacent body parts. The geometrical transformation kernels model the relationship between each pair of feature maps of adjacent parts. In order to schematically illustrate the process described above, an example is shown in FIG. 5. Feature maps 5200 and 5300 representing an right elbow and a right lower arm, respectively, are extracted from an input image 5100. One of the lower-arm feature maps 5300 has a high response, since the lower-arm feature map 5300 is extracted from a channel describing downward lower arm without clothes covered. An elbow feature map 5200, positively correlated with the lower-arm feature map 5300, also has a high response. It is expected to use the lower-arm feature map 5300 to reduce false alarms and enhance the response on the right elbow. However, it is not suitable to add the elbow feature map 5200 directly to the lower-arm feature map 5300, since there is a spatial mismatch between the two joints. Instead, the lower-arm feature map 5300 is shifted towards the right elbow through a convolution operation with a geometrical transformation kernel 5400 as described above. Afterwards, a transformed feature map 5500 is combined with the elbow feature map 5200 to generate an updated feature map 5600, leading to an improved pose estimation result.

An exemplary bi-directional message passing process is illustrated in FIG. 6. While the message is passed from the bottom to the top in the part-feature network 6100, the message is passed in an opposite direction in the part-feature network 6200. In either direction, the message is passed in a similar manner as describe above with reference to FIG. 5. Each node in the network 6100 or 6200, except for the leaf node 6105, 6106, 6109, 6110 and 6201, may be updated with the message passed from the upstream node. Taking the node 6104 for example, this process may be expressed with the following equation,

A′₄=f(A₄+A′₅⊗w^a5,a4)  (2)

wherein A′₄ represents the updated part-feature maps after message passing, A₄ represents the part-feature maps before message passing, and w^a5,a4 represents a combination of transformation kernels between the node 6105 and the node 6104. As the node 6103 may receive messages from both the node 6104 and the node 6106, the part-feature maps at node 6103 may be updated by the following equation,

A′₃=f(A₃+A′₄⊗w^a4,a3+A′₆⊗w^a6,a3).  (3)

The part-feature maps in the network 6200 may be updated in a similar way but an opposite direction, and are therefore not discussed in detail here. Finally, two sets of updated part-feature maps (A′_k, B′_k) may be linearly combined into a set of score maps indicating the likelihood of the existence of the corresponding body parts.

As will be appreciated by one skilled in the art, the present application may be embodied as a system, a method or a computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment and hardware aspects that may all generally be referred to herein as a “unit”, “circuit”, “module”, or “system”. Much of the functionality and many of the principles when implemented, are best supported with or integrated circuits (ICs), such as a digital signal processor and software therefore or application specific ICs. It is expected that one of ordinary skill, notwithstanding possibly significant effort and many design choices motivated by, for example, available time, current technology, and economic considerations, when guided by the concepts and principles disclosed herein will be readily capable of generating ICs with minimal experimentation. Therefore, in the interest of brevity and minimization of any risk of obscuring the principles and concepts according to the present application, further discussion of such software and ICs, if any, will be limited to the essentials with respect to the principles and concepts used by the preferred embodiments. In addition, the present application may take the form of an entirely software embodiment (including firmware, resident software, micro-code, etc.) or an embodiment combining software. For example, the system may comprise a memory that stores executable components and a processor, electrically coupled to the memory to execute the executable components to perform operations of the system, as discussed in reference to FIGS. 1-6. Furthermore, the present application may take the form of a computer program product embodied in any tangible medium of expression having computer-usable program code embodied in the medium.

Claims

1. A method for pose estimation, comprising:

extracting a plurality of sets of part-feature maps from an image, each set of the extracted part-feature maps representing a body part and forming a node of a part-feature network;

passing a message of each set of the extracted part-feature maps through the part-feature network to update the extracted part-feature maps, resulting in each set of the extracted part-feature maps incorporating the message of upstream nodes; and

estimating, based on the updated part-feature maps, the body part within the image.

2. The method of claim 1, wherein

the passing of the message is performed twice in opposite directions and each pairs of the updated part-feature maps performed in different directions are combined into a score map; and

the estimating of the body part is performed based on the combined score maps.

3. The method of claim 1, wherein the extracting of the part-feature maps is performed via a CNN.

4. The method of claim 3, wherein the CNN is a VGG net.

5. The method of claim 4, wherein three pooling layers are adopted in the VGG net.

6. The method of claim 1, wherein the passing of the message is performed by a convolution operation with a geometrical transformation kernel.

7. A system for pose estimation, comprising:

a memory that stores executable components; and

a processor electrically coupled to the memory to execute the executable components for: extracting a plurality of sets of part-feature maps from an image, each set of the extracted part-feature maps representing a body part and forming a node of a part-feature network; passing a message of each set of the extracted part-feature maps through the part-feature network to update the extracted part-feature maps, resulting in each set of the extracted part-feature maps incorporating the message of previously passed nodes; and estimating, based on the refined part-feature maps, the body part within the image.

8. The system of claim 7, wherein

the passing of the message is performed twice in opposite directions and each pairs of the updated part-feature maps performed in different directions are combined into a score map; and

the estimating of the body part is performed based on the combined score maps.

9. The system of claim 7, wherein the extracting of the part-feature maps is performed via a CNN.

10. The system of claim 9, wherein the CNN is a VGG net.

11. The system of claim 10, wherein three pooling layers are adopted in the VGG net.

12. The system of claim 7, wherein the passing of the message is performed by a convolution operation with a geometrical transformation kernel.