DQiaole
commented
5 days ago Sorry, the code of visualization can't be found anymore. But you can implement it according to the detials provided in the supplementary. As for the image boundaries, padding with zeros have little impact on the final results we get, as pixels within image boundaries only account for a little part of the whole datasets. By the way, you can also utilize a out-of-the-range mask before the softmax or just drop this part of pixels.
Thank you for your excellent work! I'm very interested in your research on cost volume visualization, and I've also seen your discussions with the authors of GMFlowNet. I hope you don't mind me asking if you could share your code for visualizing the cost volumes? Also, I have a question regarding the image boundaries - would cases where delta u and delta v are only available in positive or negative directions at the borders lead to unreasonable averaging results? Or how should we handle such cases? Thank you very much! Looking forward to your response!