search

maybeLx / MVSFormerPlusPlus

Codes of MVSFormer++: Revealing the Devil in Transformer’s Details for Multi-View Stereo (ICLR2024)
Apache License 2.0
180 stars 6 forks source link

Testing on Tanks and Temples #17

Open mazeyu opened 5 months ago

mazeyu commented 5 months ago

Hi, I am using the script in the MVSFormer repo to test on Tanks and Temples (training set, Ignatius, etc.) with --prob_threshold set to a scalar instead and other settings unchanged. And the resulting point cloud seems too dense, which makes the recall and precision not balanced:

evaluation result : Ignatius distance tau : 0.003 precision : 0.6671 recall : 0.9895 f-score : 0.7969

The f-score should usually be around 0.9 for Ignatius.

Is there anything setting wrong? Thanks!

ewrfcas commented 5 months ago

If you need to improve the precision, increase the prob_threshold is a good choice.

mazeyu commented 1 month ago

Hi. Could you provide the setting to reproduce T&T benchmark results (if the current prob_threshold is not what is used)? My question is more about reproducing the results instead of asking for some improvement. Thanks!