Closed DIVE128 closed 2 years ago
Great question. Actually, the MVSNet uses regression and RMVSNet uses regression. And some other methods also have different choices. In our paper, we use a binary search pipeline so we naturally use classification. The possible risk of the wrong prediction is handled by ETB. If you want more discussion about the two kinds of methods, I recommend you to read the paper of this code UniMVSNet.
Thanks for your great work. As for the depth generation, you choose argmax of classification instead of weighed operation of regression. Does it increase the risk to predict a wrong depth range in the early stage? What do you think about the choice of classification and regression in multi-view stereo.