search

princeton-vl / RAFT-3D

BSD 3-Clause "New" or "Revised" License
233 stars 26 forks source link

Question about the train_kitti #6

Closed MiaoMiaoMeng closed 3 years ago

MiaoMiaoMeng commented 3 years ago

Hi， Thanks for sharing the wonderful work. I have some question about the 'train_kitti.py', I found 'kitti.py' in your project’s folder ‘data_readers’ is for submission the kitti datasets instead of training. In train_kitti.py I can‘t found class KITTI, and in KITTIEval we can’t get object ‘flowgt’ and ‘valid’ I try to code it by my self，so I add ‘flow_gt = flowgt[self.crop:]’ and _valid = flowgt in kitty.py but I got some exception like that image

Could you upload the KITTI and KITTIEval both？

By the way, the reason why I need to retrain the model is that I cannot get the expected results from your pre-trained model. image

MiaoMiaoMeng commented 3 years ago

Okay, I think I found the problem. In raft.raft_bilaplacian, should we perform a log() operation on Ts and then return it? Because if we don’t log() it to tensor，we will get an object of 'map' type. And now I get the same result as the paper image

zachteed commented 3 years ago

Hi, I just posted the KITTI training / eval scripts so you should be able to finetune the models on the KITTI dataset now.

BillyYuZJU commented 1 year ago

@MiaoMiaoMeng Hi, I meet the same problem when try to run the demo. Could you elaborate on how to operate on Ts to get the same result as the paper?

Hi， Thanks for sharing the wonderful work. I have some question about the 'train_kitti.py', I found 'kitti.py' in your project’s folder ‘data_readers’ is for submission the kitti datasets instead of training. In train_kitti.py I can‘t found class KITTI, and in KITTIEval we can’t get object ‘flowgt’ and ‘valid’ I try to code it by my self，so I add ‘flow_gt = flowgt[self.crop:]’ and _valid = flowgt in kitty.py but I got some exception like that image

Could you upload the KITTI and KITTIEval both？

By the way, the reason why I need to retrain the model is that I cannot get the expected results from your pre-trained model. image