Open weiyithu opened 2 years ago
I use the DGP data load to load the DDAD six camera at the same time, tqdm show it has 12k, I guess it is because the DGP will discard some small action frames or static frames, the DGP code is so complicated, so I have not read carefully, I just use it.
Thanks for your responce! I think 12k (12650) is right but total number of images should not be 63250. By the way, have you successfully reproduced the results ?
I apologize if there is any confusion. We are preparing an updated version of the paper and will upload it to arXiv soon, and we have started looking into open-sourcing the FSM code.
Hi author, do you use the depth from DDAD dataloader with original resolution (1216 x 1936) for evaluation?
I apologize if there is any confusion. We are preparing an updated version of the paper and will upload it to arXiv soon, and we have started looking into open-sourcing the FSM code.
No, I have also not reproduced the result. I have tried many setups but its RMSE is still worse than the paper 5, so I give up now...
I apologize if there is any confusion. We are preparing an updated version of the paper and will upload it to arXiv soon, and we have started looking into open-sourcing the FSM code.
Hello, I am confused that do you using automask when calculation photometric loss on FSM? The description of that on the paper is not clear.
I apologize if there is any confusion. We are preparing an updated version of the paper and will upload it to arXiv soon, and we have started looking into open-sourcing the FSM code.
Hello, I am confused that do you using automask when calculation photometric loss on FSM? The description of that on the paper is not clear.
Yeah, we annotate and use self-occlusion mask in DDAD.
Thanks for your great work. I have some questions about the evaluation on DDAD. For the validation set, DDAD has 3950 samples, we should get 39506=23700 images but not 15800 images. Also for the training set, we should get 126506=75900 images but not 63250 images.