a question about evaluate_FlowFormer.py

[deqings]

Very interesting work!

I have a question about evaluate_FlowFormer.py, which reports AEPE 0.63/1.51 on Sintel clean/final for the provided things.pth checkpoint.
flow_pre = model(image1, image2)
flow_pre = padder.unpad(flow_pre[0]).cpu()
epe = torch.sum((flow_pre - flow_gt)**2, dim=0).sqrt()
For my run on Sintel images, flow_pre is of size [1, 2, 436, 1024] and flow_gt [2, 436, 1024]. (flow_pre - flow_gt)**2 is of size [1, 2, 436, 1024] and sum over dim 0 results in size [2, 436, 1024] for epe. Thus, "epe" doesn't sum over the horizontal and vertical flow errors.
flow_pre = torch.squeeze(padder.unpad(flow_pre[0]).cpu())
would results in size [436,1024] for epe and reports AEPE 1.01/2.40.

Could you please help check it?

Thanks.

[drinkingcoder]

Thanks for pointing that out. Sorry for the mistake 😵. We will roll back our Arxiv paper to the ECCV submission version, where the reported AEPE is 0.95/2.35, calculated through the tile technique.

[deqings]

Thanks for checking it. These are still impressive numbers.

BTW: if you haven't submit your new ArXiv paper, could you please also correct the three A+S+K+H entries in Table 1? The numbers there are for the training set of Sintel and KITTI after the models have been trained on AutoFlow. Thanks.

[drinkingcoder]

Many many thanks for your reminding! We will update that!