Reproduce “JRDB-Pose: A Large-scale Dataset for Multi-Person Pose Estimation and Tracking” on OpenResearch

About this paper

Edward Vendrow, Duy Tho Le, Jianfei Cai, Hamid Rezatofighi

JRDB-Pose is introduced as a large-scale, high-frequency dataset for multi-person pose estimation and tracking, featuring 636,000 pose instances and 11 million labeled keypoints collected from a social navigation robot in crowded human environments. This dataset, which includes consistent tracking IDs and per-keypoint occlusion labels, presents greater challenges for state-of-the-art models compared to existing benchmarks.

Key insights

• JRDB-Pose introduces new and more difficult challenges for state-of-the-art multi-person pose estimation and tracking methods, reflected by comparatively lower performance scores (e.g., AP, MOTA) on this dataset.
• Occlusion significantly impacts model performance, with per-keypoint average errors being substantially higher for occluded and invisible joints across all evaluated methods.
• Training or finetuning models directly on JRDB-Pose is critical for achieving robust performance in robot-centric environments, as models solely pretrained on datasets like COCO without finetuning show degraded tracking capabilities due to a domain gap.

Summary by alphaXiv.