Deep Learning based Human Pose Estimation
- Publication Type:
- Thesis
- Issue Date:
- 2020
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Human pose estimation is an important research area in vision-based human activity analysis. Human pose estimation aims to estimate the human articulate joint positions in 2D/3D space from images or videos. Due to the complexity of the real environment and the diversity of human poses, vision-based human pose estimation is challenging. Recently, the rapid development of deep learning has much promoted the simulation of the analysis and reasoning capabilities of the human visual system. Therefore, it is of considerable significance to further explore vision-based human pose estimation using deep learning techniques. Specifically, this thesis proposes a series of methods for human pose estimation, summarized as follows:
We propose a video-based 2D pose estimation model, which embeds a multi-scale TCE module into the encoder-decoder network architecture for explicitly exploring temporal consistency in videos. The TCE module applies the learnable offset field to capture the geometric transformation between adjacent frames at the feature level. In addition, we explore the multi-scale geometric transformations at the feature level by integrating the spatial pyramid within the TCE module, which achieves further performance improvements.
We propose a self-supervised approach for 3D human pose estimation, which only relies on geometric prior knowledge and does not require any 3D human pose annotations. To this end, we design the transform re-projection loss, which is an effective technique to exploit multi-view consistency information and constrain the estimated 3D poses during training. Besides, we introduce a root position regression branch to restore the global 3D poses during training. In this way, the network can reserve the scale information of re-projected 2D poses, which can improve the accuracy of the predicted 3D poses.
We propose a self-supervised 3D human pose estimation method based on the consistent factorization network, which fully disentangles the 3D human shape and camera viewpoint to overcome the projection ambiguity problem. To this end, we design a simple and effective loss function using multi-view information to constrain the canonical 3D human pose. Moreover, in order to reconstruct robust canonical 3D human poses, we represent 3D human pose as a combination of a dictionary of 3D pose basis, and adopt geometric information of 3D human poses to learn a hierarchical dictionary from 2D human poses by solving the NRSfM problem.
Please use this identifier to cite or link to this item: