[深度学习论文笔记][Human Pose Estimation] DeepPose: Human Pose Estimation via Deep Neural Networks

DeepPose: Human Pose Estimation via Deep Neural Networks
Toshev, Alexander, and Christian Szegedy. “Deeppose: Human pose estimation via deep neural networks.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2014. (Citations: 271).

1 Motivation

We represent a person by K joints, and we use a CNN-based regression model towards body joints.

2 Method

2.1 Dataset

The dataset is 

Where

y^i_k contains the height and width coordinates of the k-th joint.

2.2 Normalization

Since the joint coordinates are in absolute image coordinates, it proves beneficial to normalize them wrt a box B bounding the human body or parts of it. Such a box is defined by its center b_C as well as height B_H and width B_W . 

The normalized pose vector is

The normalized image X^(i) is a crop of the image by the bounding box B.

2.3 Architecture
See Fig. This is modified from AlexNet. The input is normalized image, the output is K normalized pose vectors.

2.4 Loss
Use l_2 loss for each joint.

3 Cascade of Pose Regressors
Due to its fixed input size of 220 × 220, the network has limited capacity to look at detail. It learns filters capturing pose properties at coarse scale. These are necessary to estimate rough pose but insufficient to always precisely localize the body joints. Therefore, we propose to train a cascade of pose regressors. At the first stage, the cascade starts off by estimating an initial pose. At subsequent stages, additional CNN regressors are trained to predict a displacement of the joint locations from previous stage to the true location. See Fig. The network architecture for all stages of the cascade are the same, but they learn different network parameters.

4 Result
See Fig.