search

YoYo000 / MVSNet

MVSNet (ECCV2018) & R-MVSNet (CVPR2019)
MIT License
1.39k stars 303 forks source link

Low Depth Accuracy in Training on ScanNet #87

Open fuy34 opened 4 years ago

fuy34 commented 4 years ago

Hi, Thanks for sharing the code. I am recently trying to train the R-MVSNet on a subset of ScanNet dataset. I first run the colmap and then using the provided code to get the cam.txt and pari.txt, and the ground truth is from ScanNet directly.

During the training, I use the ground truth depth range of reference frame with max_d=128, and scale both the image and camera intrinsic by 0.25. No augmentation or refinement is used.

There is no problem to run the pipeline, but the training accuracy seems very low (attached below). Compared to DTU and ETH3D, this dataset does include more blur frames and resolution the resolution is lower (640*480). But I am not sure if this will lead to such a significant performance drop.

May I ask if you have any insight about it? Or is there any detail I have missed?

Screen Shot 2020-04-11 at 11 21 00 AM
guoyaoy commented 2 years ago

Hi, have you solved it? I trained the ScanNet too, but failed. The depth range of ScanNet(gt) among 400~3000, but the depth_min computed from colmap2mvsnet only 0.0×. It has different scales. If I want to train other datasets, how could I solve it? Thanks!