Walking is Matter: A Benchmark for Fine-Grained Gait Segmentation

Zhongguang Zhang, Wenzhu Xu, Min Tang, Yulin Zhou, Qifei Zhang, Chao Wu, Zhao Wang

Abstract: Walking is a common activity of daily living that engages multiple levels of the nervous system, various components of the musculoskeletal framework, and the cardiorespiratory system. Clinical trials need to specify specific gait characteristics where fine-grained gait analysis technique is in demand. Fine-grained gait segmentation aims to segments a temporally untrimmed gait video sequence over time with predefined gait phase labels. It is able to capture intricate details and nuances in human gait, leading to a deeper understanding of biomechanics, health diagnostics, and behavioral analysis. Comparing with closely related temporal action location techniques, few efforts have been taken on fine-grained gait segmentation. In addition, academia faces limitations imposed by existing databases that are predominantly recorded and labeled at a coarse sequence level, e.g., one action label for a motion sequence. Hence, it is essential to build a benchmark dataset for fine-grained gait segmentation purpose. To address this issue, we conduct a comprehensive benchmark study for better practicality in fine-grained gait segmentation, rather than focusing on a particular model for incremental performance. To this end, a gait dataset named FineGait is developed. Extensive gait segmentation experiments with comprehensive baselines are conducted. Results indicate that the proposed FineGait is necessary and effective for gait segmentation. We believe that the availability of a diverse and well-annotated FineGait dataset could play a pivotal role in advancing research and development in gait analysis.