lez-s
Impressive work
Thank you for the effort you have put into this excellent work. However, I noticed that this paper primarily discusses image quality metrics. Have you tried using the synthetic data for stereo matching training? Or, do you think the model is now stable and accurate enough to achieve pixel-level accuracy in non-occluded regions?
hi, I tried using KITTI dataset for training with some network modifications, but the results are worse than current methods (described in supplementary of the paper). I think you may need some techniques to maintain the consistency of the left and right images at the pixel level during training and inference. I don't think it is accurate enough, this is a simple and cost-effective method since it does not require retraining.
