Never Forget the Basics: In-distribution Knowledge Retention for Continual Test-time Adaptation in Human Motion Prediction
Keywords: Human Pose Prediction, Domain Adaptation, Graph Representation Learning, Graph Out-of-distribution Detection
Abstract: This paper presents a novel approach to addressing the underexplored challenge of human pose prediction in dynamic target domains that simultaneously contain in-distribution (ID) and out-of-distribution (OOD) data. Existing test-time adaptation (TTA) techniques predominantly focus on OOD data, neglecting the fact that ID data, which closely resembles the training distribution, is often encountered during real-world deployment, leading to significant degradation in ID performance. To address this, we introduce In-Distribution Knowledge Retention (IDKR), a continual TTA framework designed to preserve critical knowledge about ID data while adapting to unseen OOD sequences. Our method introduces an ID-informative subgraph learning strategy that leverages the structural characteristics of human skeletal data to compute a structural graph Fisher Information Matrix (SG-FIM). Unlike prior work, IDKR simultaneously considers both node and edge features in the skeletal graph, with edge features, representing the invariant bone lengths between parent-child joint pairs, being essential for maintaining structural consistency across poses. These edge features are key to extracting reliable SG-FIM parameters, enabling the model to retain parameters critical for ID performance while selectively updating those needed for OOD adaptation. Extensive experiments on multiple benchmark datasets demonstrate that IDKR consistently outperforms state-of-the-art methods, particularly in scenarios involving mixed ID and OOD data, setting a new standard for robust human pose prediction in dynamic environments.
Primary Area: applications to robotics, autonomy, planning
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: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
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.
Submission Number: 1491
Loading