Motion-R$^{3}$: Fast and Accurate Motion Annotation via Representation-based Representativeness Ranking

Chengxu Zuo, Jubo Yu, Tianxiang Ren, Shihui Guo, Andreas Aristidou, Zijiao Zeng, Yazhan Zhang, Anjun Chen, Yipeng Qin

Keywords: Motion Annotation, Representation Learning, Contrastive Learning, Motion Representation

TL;DR: A small amount of highly representative data is automatically selected for manual annotation to train an annotation model, which can then assist in annotating remaining data and reduce labor costs.

Abstract: As motion capture data collection becomes more accessible, efficient motion annotation tools are increasingly needed to streamline dataset labeling. In this paper, we propose a data-centric motion annotation method that leverages the inherent representativeness of motion data. Specifically, we introduce Representation-based Representativeness Ranking (R3), which ranks motion samples based on their significance in a learned representation space. To enhance this space, we develop a dual-level motion contrastive learning approach, improving the informativeness of the learned representations. Our method is designed for high efficiency and adaptability, making it particularly responsive to frequent requirement changes. By employing unsupervised contrastive pre-training, we reduce the labeling time for expert oracles while using a lightweight classifier to accelerate annotation predictions. Additionally, we incorporate active learning to recommend more representative data, minimizing the number of required expert annotations while maintaining annotation quality. Experimental results demonstrate that our approach outperforms state-of-the-art methods in both accuracy and efficiency, enabling agile development of motion annotation models.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 2994