Lahav Lipson

With some modifications it can support right->left disparity. The correlation sampler returned by `corr_block()` assumes left->right disparity so the order of the input feature maps will need to be flipped....

EPE is just the euclidean distance between the predicted flow/disparity vector and the GT flow/disparity vector. Typically methods report the average EPE over the entire image. The equation above seems...

You are correct, the `dim = 0` is not needed. This evaluation code is meant to also work for 2D flow. The `>= 0.5` is to convert an array of...

Here is one ``` >>> torch.tensor([1,0,1]) >= 0.5 tensor([ True, False, True]) ```

I can't reproduce this issue. From the root directory, I can obtain the file information for a Middlebury GT using the following command: ``` (raftstereo) llipson@node803 RAFT-Stereo$ identify datasets/Middlebury/MiddEval3/trainingF/PianoL/disp0GT.pfm datasets/Middlebury/MiddEval3/trainingF/PianoL/disp0GT.pfm...

Are you setting `test_mode=True` in the forward pass? I get the same error in this case.

GatherElements is a built-in Pytorch function, not something our code implements.

These results you generated are the same that were submitted to the middlebury training dataset scoreboard. The difference is in the evaluation, which I believe prioritizes difficult image regions like...

All functions in [evaluate_stereo.py](https://github.com/princeton-vl/RAFT-Stereo/blob/main/evaluate_stereo.py) report threshold error metrics.

We have three known equations: inverse projection from (2D->3D): ((u0 - cx0)/fx) * z = X0 camera translation (3D->3D): X1 = X0 + baseline projection from (3D->2D) u1 = fx*(X1/z)...