ACTUPose: Active Curriculum Training for Unsupervised Domain Adaptation in Pose Estimation

Isha Dua, Arjun Sharma, Shuaib Ahmed, Rahul Tallamraju

Keywords: Unsupervised Domain Adaptation, Synthetic Data, Domain Adaptation, Human Pose Estimation

TL;DR: ACTUPose is a training strategy that utilizes the diversity of poses in the source data for effective domain adaptation.

Abstract: In this study, we present a novel training methodology for unsupervised domain adaptation (UDA) in the context of pose estimation. Existing UDA methods for pose estimation often struggle to generalize effectively across similar poses in the target data, even when such poses are well represented in the source data. We attribute this challenge to the lack of uniform support for different poses and a systematic training strategy for handling poses of varying complexity in the source domain. To address this challenge, we propose ACTUPose, an active curriculum training strategy that utilizes the diversity of poses in the source data. Within this framework we incorporate a method for quantifying pose complexity that dynamically orders the training data as the training progresses. We further introduce a new loss function aimed at improving skeletal structure prediction. Additionally, we incorporate a cross-domain feature loss to better utilize unlabeled real data. With this approach we demonstrate state-of-the-art performance across standard benchmark datasets for UDA in Pose Estimation.

Submission Number: 65