search

jzhangbs / Vis-MVSNet

Visibility-aware Multi-view Stereo Network
MIT License
235 stars 27 forks source link

Is the pretrained model 20000.tar is trained on DTU dataset or the BlendedMVS dataset? #32

Open VILmchen opened 2 years ago

VILmchen commented 2 years ago

Thanks for your nice work! Is the pre-trained model 20000.tar is trained on the DTU dataset or the BlendedMVS dataset? I directly use the pre-trained model 20000.tar and the sh/tnt.sh script and the fusion.py only get an F-score 55.97 on Tanks and Temples, is there any hyper-parameters that need to change?

jzhangbs commented 2 years ago

Hi, are you using the provided preprocessed TnT test set with provided sfm result instead of the one from colmap? Current fusion script can achieve 58.41 with the same setting.

VILmchen commented 2 years ago

Thanks for your reply. I use the preprocessed TnT test set which is downloaded from the repo of YoYo000/MVSNet. Besides, I want to confirm that is the pre-trained model 20000.tar is trained on the DTU dataset or the BlendedMVS dataset?

jzhangbs commented 2 years ago

it's trained on BlendedMVS

VILmchen commented 2 years ago

OK, thanks very much! Is there any methods to obtain the results shown on the leaderboard of TnT by the fusion script?

jzhangbs commented 2 years ago

We are working on a better re-implementation of the fusion code. We will release it soon.

VILmchen commented 2 years ago

OK, thanks!