Domain Adaptation for pose estimation in low quality images (Abstract)

This paper addresses a domain adaptation model for estimating human pose in low quality images without label information. Given labeled high quality images (source domain) and unlabelled low quality images (target domain), we propose a Latent Self-Adaptive Support Vector Machine (LSASVM) method to adapt the existing pose estimation model for high quality images to the one for low quality images. To solve the problem of no labeled data in low quality images, we also propose Quality-Guided Transfer (QGT) method to generate the data in low quality images based on the quality information in two domains. And a latent model is utilized to measure the adapted degree of each body part. Although the results of our preliminary experiment are not good enough, the performance of some images increases by our approach.