General Skeleton Semantics Learning with Probabilistic Masked Context Reconstruction for Skeleton-Based Person Re-Identification

Haocong Rao, Chunyan Miao

Keywords: General skeleton semantics learning, Generality Assessment, Skeleton-based person re-identification, Probabilistic masked reconstruction, Spatial-temporal context learning

Abstract:

Person re-identification (re-ID) via skeleton data is an emerging topic with immense potential for safety-critical applications. Existing methods usually utilize spatial or temporal skeleton semantics learning (SSL) tasks to facilitate skeleton representation learning, while most SSL tasks are *model-dependent* and lack the ability to capture general fine-grained (*e.g.*, joint-level) spatial-temporal skeleton patterns under different model architectures. To delve into multi-faceted generality of SSL tasks, we first propose an SSL generality assessment framework termed **SCUT** that identifies four key SSL properties: **S**patial-temporal effectiveness, **C**o-training compatibility, **U**nsupervised trainability, and **T**ask transformability. By formulating systematic evaluation criteria for each property, SCUT enables both qualitative and quantitative analysis of SSL generality under varying models and scenarios. Motivated by SCUT to fully harness skeleton context for semantics learning, we further devise a generic **Pro**babilistic **M**asked S**p**atial-**T**emporal cont**e**xt **R**econstruction (**Prompter**) task to enhance performance of skeleton-based person re-ID models. Specifically, Prompter first probabilistically and independently masks joints' structural locations to generate *spatial context*, and then randomly conceal their motion trajectories to form *temporal context*.
Through combining both spatial and temporal skeleton context representations to jointly reconstruct and infer skeleton sequences, Prompter encourages the model to capture general valuable spatial-temporal skeleton patterns for person re-ID. Empirical evaluations on SCUT and five benchmark datasets demonstrate the superiority of Prompter to most state-of-the-art SSL tasks. We further validate its general effectiveness in different skeleton modeling, RGB-estimated or cross-domain scenarios

Primary Area: applications to computer vision, audio, language, and other modalities

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: 9273