# HumanGPS: Geodesic PreServing Feature for Dense Human Correspondences Source: [HumanGPS: Geodesic PreServing Feature for Dense Human Correspondences]() Collection index: [llms.txt]() Data updated: 2025-12-24T15:38:39.854Z Authors: Feitong Tan, Danhang Tang, Mingsong Dou, Kaiwen Guo, Rohit Pandey, Cem Keskin, Ruofei Du, Deqing Sun, Sofien Bouaziz, Sean Fanello, Ping Tan, Yinda Zhang Published: 2021-03-29T12:43:44.000Z ## Abstract In this paper, we address the problem of building dense correspondences between human images under arbitrary camera viewpoints and body poses. Prior art either assumes small motion between frames or relies on local descriptors, which cannot handle large motion or visually ambiguous body parts, e.g., left vs. right hand. In contrast, we propose a deep learning framework that maps each pixel to a feature space, where the feature distances reflect the geodesic distances among pixels as if they were projected onto the surface of a 3D human scan. To this end, we introduce novel loss functions to push features apart according to their geodesic distances on the surface. Without any semantic annotation, the proposed embeddings automatically learn to differentiate visually similar parts and align different subjects into an unified feature space. Extensive experiments show that the learned embeddings can produce accurate correspondences between images with remarkable generalization capabilities on both intra and inter subjects. ## Recorded benchmark results A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage. - [Dense Correspondence Search on SMPL Intra-Subject (test)](): Dense Correspondence Search; SMPL Intra-Subject (test); EPE (non-occluded): 7.12. - [Dense Correspondence Search on Relightables Intra-Subject (test)](): Dense Correspondence Search; Relightables Intra-Subject (test); EPE (Non-Occluded): 11.24. - [Dense Correspondence Search on RenderPeople Intra-Subject (test)](): Dense Correspondence Search; RenderPeople Intra-Subject (test); EPE (Non-Occluded): 11.91. - [Dense Correspondence Search on SMPL Inter-Subject (test)](): Dense Correspondence Search; SMPL Inter-Subject (test); EPE (non-occluded): 8.49. - [Dense human correspondences on SMPL (Intra-Subject)](): Dense human correspondences; SMPL (Intra-Subject); Non-Correspondence Metric: 2.13. - [Dense human correspondences on Relightables (Intra-Subject)](): Dense human correspondences; Relightables (Intra-Subject); Non-Correspondence Error Rate: 2.27. - [Dense human correspondences on RenderPeople Intra-Subject](): Dense human correspondences; RenderPeople Intra-Subject; Non-MPJPE: 3.95. - [Occlusion Detection on SMPL Intra-Subject (test)](): Occlusion Detection; SMPL Intra-Subject (test); Average Precision: 94.93. - [Occlusion Detection on The Relightables Intra-Subject (test)](): Occlusion Detection; The Relightables Intra-Subject (test); AP: 87.67. - [Occlusion Detection on RenderPeople Intra-Subject (test)](): Occlusion Detection; RenderPeople Intra-Subject (test); AP: 91.38. - [Occlusion Detection on SMPL Inter-Subject (test)](): Occlusion Detection; SMPL Inter-Subject (test); AP: 80.91. - [Optical Flow Estimation on Single-Human Optical Flow (SHOF)](): Optical Flow Estimation; Single-Human Optical Flow (SHOF); AEPE: 0.1239. - [Dense Human Pose Regression on DensePose MSCOCO (test)](): Dense human pose regression; DensePose MSCOCO (test); Error (5cm Threshold): 55.41. - [Temporal Consistency on Synthetic Dataset 18,000 frames sequence](): Temporal Consistency; Synthetic Dataset 18,000 frames sequence; PCC (Interval 1): 86.42. - [Dense human correspondences on SMPL (Inter-Subject)](): Dense human correspondences; SMPL (Inter-Subject); Non-Correspondence Rate: 12.08. ## Original sources and implementation [Original paper]() [PDF]() [Implementation]()