Masked Representation Learning for Domain Generalized Stereo Matching

Overview Contributions Ø We build a pseudo-multi-task learning framework to increase generalization. Ø Our methods can improve cross-domain accuracy and reduce the volatility. Ø We find that cross-domain results varies significantly among different epochs. Experiments Cross-domain generalization evaluation (peak results) on four target datasets. The fine-tuning results on the KITTI dataset. Conclusion: Our method can help model improve cross-domain performance, but it seems no help for fine-tuning.