Animal Pose Refinement in 2D Images with 3D Constraints
Xiaowei Dai (Sichuan University),* Shuiwang Li (Guilin University of Technology), Qijun Zhao (Sichuan University), hongyu yang (sichuan university)
The 33rd British Machine Vision Conference
Abstract
Animal pose has many potential applications in various fields. However, uncontrollable illumination, complex backgrounds and random occlusions in in-the-wild animal images often lead to large errors in pose estimation. To address this problem, we propose a method for refining the initial animal pose with 3D prior constraints. First, we learn a 3D pose dictionary from synthetic data with each atom providing 3D pose prior knowledge. Then, the 3D pose dictionary is used to linearly represent the potential 3D pose corresponding to the 2D pose that has been initially estimated for the animal in 2D image. Finally, the representation coefficients are optimized to minimize the difference between the initially-estimated 2D pose and the 2D-projection of the potential 3D pose. Moreover, to deal with the data scarcity, we construct 2D and 3D animal pose datasets, which are used to evaluate algorithm performance and learn 3D pose dictionary, respectively. Experimental results show that the proposed method is capable to utilize 3D pose knowledge well and is effective in improving 2D animal pose estimation.Video
Citation
@inproceedings{Dai_2022_BMVC,
author = {Xiaowei Dai and Shuiwang Li and Qijun Zhao and hongyu yang},
title = {Animal Pose Refinement in 2D Images with 3D Constraints},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0848.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by