Training from scratch problem about sc-sfmlearner-release HOT 8 CLOSED

1. Overfitted on KITTI? I tried many times but didn't meet this issue. I may need more info to reproduce.
2. Yes. GC loss and smoothness loss need some time to converge to correct ranges.

from sc-sfmlearner-release.

Yes, I trained it on KITTI.

from sc-sfmlearner-release.

Do you use GT for validation? What's the metric for the training loss and validation loss in the first figure? Is it AbsRel or photometric+GC+smooth?

from sc-sfmlearner-release.

I run the default train_resnet50_pose_256.sh. There is no --with-gt in the default scripts. Fig1 is the curve of the log progress_log_summary_csv.

from sc-sfmlearner-release.

I see. I didn't provide Gt for KITTI Odometry dataset. Actually, the unsupervised loss (photometric+GC+smoothness) is not good enough for validation purpose. You just need train the model for more than about 50 epoches. Then the 'best' or the 'latest' models are both good enough.

from sc-sfmlearner-release.

Thanks for your reply. Did you change the learning rate during training?

from sc-sfmlearner-release.

Never. I find that it doesn't matter. However, I suggest training from ImageNet pretrained model. It would be much better, and it is the main reason why I update network in this version. ResNet pretrained models are publicly available. So why not use.

from sc-sfmlearner-release.

Thanks a lot. I'll try.

from sc-sfmlearner-release.