FlexPose: Pose Distribution Adaptation with Few-shot Guidance

Zixiao Wang, Junwu Weng, Yong Zhang, Shaoli Huang, Xuan Wang, Jue Wang, Bei Yu

Keywords: Pose Adapation, Human Pose Detection, Few-shot

TL;DR: We transfer a human pose distribution to another one with only few-shot guidance and apply it to multiple pose-based tasks.

Abstract: Annotating human pose images can be costly. Meanwhile, there is an unavoidable major performance drop when a pre-trained pose estimation model is evaluated on a new dataset. We observe that pose distributions from different datasets share similar pose priors with different geometric transformations, which inspires us to learn a pose generator that can flexibly be adapted to generate the pose of a new pose distribution. In this paper, we treat human poses as skeleton images and propose a scheme to transfer a pre-trained pose annotation generator to generate poses from the transferred distribution of a newly collected dataset with only a few annotation guidances. By finetuning a limited number of linear layers, the transferred generator is able to generate similar pose annotations to the target pose distribution. We evaluate our FlexPose on several cross-dataset settings qualitatively and quantitatively. FlexPose surprisingly achieves around 41.8$\%$ performance improvement on the Unsupervised Pose Estimation task when it transfers the pose distribution of COCO, 3DHP and Surreal dataset to that of the H36M dataset.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)