Why do we need InclusiveVidPose?

Approximately 40 million individuals worldwide live with limb deficiencies. These people could benefit greatly from accurate human pose estimation (HPE) in applications such as rehabilitation monitoring and health assessment. Regrettably, existing HPE systems trained on standard datasets fail to accommodate atypical anatomies and prosthetic occlusions, leaving this population unsupported. However, widely used benchmarks like MS COCO and MPII Human Pose include only able-bodied people with complete sets of keypoints. These datasets and the methods built on them assume that every keypoint of a presented individual exists, making no provision for missing or altered limbs. As demonstrated in the figure next, the MS COCO–trained ViTPose model produces significant errors when applied to images of individuals with limb deficiencies. Consequently, people with limb deficiencies are excluded from current benchmarks, and models trained on these datasets fail to generalize to their anatomies. To address this gap, we introduce InclusiveVidPose, the first video-based HPE dataset focused on individuals with limb deficiencies.