hmorimitsu/ptlflow

results of raft on sintel after training on kitti

Closed this issue · 2 comments

Hi Henrique, thanks for this great repo! I have a question about the evaluation results of raft on sintel.
The results on sintel are 6.293/4.687 after training on kitti. On the RAFT paper the authors reported 1.61/2.86 (Table 1).
I understand the model was fine-tuned on kitti only after training on sintel. Therefore the model performance on sintel drops. I also obtained similar results after this training procedure.

In order to achieve the results reported in the RAFT paper, seems we should mix the sintel data with KITTI. ("When evaluating on the Sintel(test) set, we finetune on the combined clean and final passes of the training set along with KITTI and HD1K data. ") Do you have plan to do so? Thanks.

Hi @askerlee, thanks for the interest. Unfortunately, right now I do not have time to train that C+T+S+K+H (Table 1) model by myself. At the moment, the weights for RAFT are just directly adapted from the original ones.

I think there is no description about whether the provided weights are exactly the ones used in the paper. However, if I had to guess, I would say that my raft-kitti weights are obtained by the C+T+K training. And it looks like that my raft-sintel weights could be those from the C+T+S+K+H training, but I am not sure. You could try asking the authors about the weights, maybe they can provide more info about them.

I might do my own training at some time in the future, but I still do not know when I will be able to do it.

Thanks. I've asked Zach and he said the sintel stage training of RAFT is actually mixing Sintel, KITTI and HD1K. Sorry for my confusion.