Abstract

The inconsistency between training strategies and task objectives is a principal constraint impeding DensePose development. DensePose aims to minimize the surface distance between the estimations and the ground truth, thereby establishing a correspondence between two-dimensional images and threedimensional humans. However, the discrete IUV optimization strategy employed by the existing pipeline deviates from the goal of minimizing surface distance, thus yielding sub-optimal results. To solve this problem, we propose Geodesic-consistent RCNN (GC RCNN), which models the intrinsic interdependence within the IUV to enhance human surface understanding and facilitate surface distance optimization. GC RCNN incorporates the adaptively semantic enhancement module (ASEM) and the geodesic-consistent loss (GCL) into the traditional DensePose estimation framework. ASME enhances human surface semantics by extracting distinct IUV features and dynamically modeling the feature-level IUV interdependence. GCL extracts the surface distance information by quantifying the degree of deviation between the estimated IUV distribution and the ground truth. The distribution deviation is utilized to refine the IUV optimization process, thereby aiding in minimizing surface distance. Empirical analyses conducted on the DensePose-COCO dataset validate the superior performance of GC RCNN, which surpasses the base model by a margin of up to 3.3% AP.