As a dense prediction task aimed at instance-level human analysis, dense-pose estimation seeks to accurately map 2D pixels onto the 3D surface of the human body. Despite significant progress has been made, two major challenges continue to confront the research community: the first is training instability caused by a large number of surface points to be regressed; the second is the significant amount of time and computational resources to manually adjust multi-task loss weights. To overcome these challenges, we present a novel dense pose estimator, named UV R-CNN, which is based on a detailed analysis of the loss formulation used in existing algorithms. The proposed UV R-CNN first introduces a novel surface point regression loss, which serves to constrain the immense loss and stable the training progress, named Dense Points Loss (DP-Loss). Additionally, we incorporates a Balanced Weighting Strategy (BWS) that allows for the automatic adaptation of loss weights. Remarkably, without auxiliary supervision and external knowledge from other tasks, UV R-CNN can be trained with larger learning rate, achieving 65.0% AP_gps and 66.1% on the DensePose-COCO validation subset with ResNet-50-FPN as backbone, competitive to the state-of-the-art methods.