SRC-Disp: Synthetic-Realistic Collaborative Disparity Learning for Stereo Matching

摘要

Stereo matching task has been greatly improved by convolutional neural networks, especially the fully-convolutional network. However, existing deep learning methods always overfit to specific domains. In this paper, focus on domain adaptation problem of disparity estimation, we present a novel training strategy to conduct synthetic-realistic collaborative learning. At first, we design a compact model that consists of shallow feature extractor, correlation feature aggregator and disparity encoder-decoder. Our model enables end-to-end disparity regression with fast, speed and high accuracy. To perform collaborative learning, we then propose two distinct training schemes, including guided label distillation and semi-supervised regularization, both of which are used to alleviate the lack of disparity labels in realistic datasets. Finally, we evaluate the trained models on different datasets that belong to various domains. Comparative results demonstrate the capability of our designed model and the effectiveness of collaborative training strategy.