Cross-Domain Complementary Learning Using Pose for Multi-Person Part Segmentation

Supervised deep learning with pixel-wise training labels has great successes on multi-person part segmentation. However, data labeling at pixel-level is very expensive. To solve the problem, people have been exploring to use synthetic avoid labeling. Although it easy generate for data, results are much worse compared those using real and manual The degradation of performance mainly due domain gap, i.e., discrepancy pixel value statistics between data. In this paper, we observe that humans both a skeleton (pose) representation. We found skeletons can effectively bridge domains during training. Our proposed approach takes advantage rich realistic variations easily obtainable learn segmentation images without any human-annotated labels. Through experiments, show human labeling, our method performs comparably several state-of-the-art approaches which require Pascal-Person-Parts COCO-DensePose datasets. On other hand, if also available in real-images training, outperforms supervised methods by large margin. further demonstrate generalizability predicting novel keypoints where no detection. Code pre-trained models https://github.com/kevinlin311tw/CDCL-human-part-segmentation.