PSMNet
PSMNet copied to clipboard
JiaRenChang
can not reach your results of val error 1.83 on KITTI15
Hi @JiaRenChang I pretrained your stackhourglass network on sceneflow dataset for 10 epoch, and then finetune it on KITTI15 with your provided train/val split for another 300 epoches. However, I can only get 2.8% error rate which is far from your result of 1.83%. All settings are the default ones provided in your code. Did I miss anything? Looking forward to your reply.
@shanyuhu Did you check the validation set are those ones? [1,3,6,20,26,35,38,41,43,44,49,60,67,70,81,84,89,97,109,119,122,123,129,130,132,134,141,144,152,158,159,165,171,174,179,182, 184, 186,187,196]
Yes, I use the same validation set as yours. One difference is that I use larger batch sizes instead of 12 during training. Do you think this will be the reason?
By the way, I also use your provided kitti_15 model to evaluate on KITTI15_test. The results are slightly worse than yours in the KITTI leaderboard. don't know why?
@shanyuhu We tried batch size = 32,16,12,8,4 in our experiment. We found that the best performance on training scene flow is using batch size = 8 or 12.
It may related to that pytorch updated their upsample function. I am not sure.
@JiaRenChang thanks for your reply. Btw, whether the provided kitti15 model is the one you used for the final kitti15 leaderboard submission?
@JiaRenChang Is the feature extractor in the sceneflow model pre-trained using other dataset, like imagenet?
@shanyuhu Did you check the validation set are those ones? [1,3,6,20,26,35,38,41,43,44,49,60,67,70,81,84,89,97,109,119,122,123,129,130,132,134,141,144,152,158,159,165,171,174,179,182, 184, 186,187,196]
you refer to indexes in [0,199] or [1, 200]? I suppose the first, am I correct?
Could you share your finetune PSMNet with me? As my computer's memory is not enough?
Can you tell me how to divide the training set and validate set on KITTI2012, preferably the picture number, thank you!
@Chenlei6 kitti 2012 val set [3,6,20,26,38,41,43,44,49,60,67,70,81,84,89,97,109,119,122,123,129,130,132,134,141,144,152,158,165,171,174,179,184,186]
I‘m sorry to disturb you! Can you share the training method for getting the best model on the KITTI2015 dataset? In the paper , you said that you have trained 1000 epochs. Is there any other training strategy to share?