search

prstrive / UniMVSNet

[CVPR 2022] Rethinking Depth Estimation for Multi-View Stereo: A Unified Representation
MIT License
231 stars 12 forks source link

The result on ETH3D dataset #4

Closed CanCanZeng closed 2 years ago

CanCanZeng commented 2 years ago

Hi, thank you for opening this great project! I tested this algorithm on Tandks and Temples dataset, the result is really nice! But when I tried on ETH3D dataset, the result is not so good, have you tested on ETH3D dataset? The checkpoint I used is the provided model on Tanks and Temples.

prstrive commented 2 years ago

Sorry, we didn't experiment with ETH3D. We think the reason for the suboptimal should be related to the point cloud filtering and fusion configuration, which you can tune appropriately according to the latest paper.

CanCanZeng commented 2 years ago

Hi, I tried different fusion strategy (ACMP and COLAMP), and truned fusion parameters in UniMVSNet, but all of these methods could not produce very nice result like in Tanks and Temples dataset. I mean, the result is still good, but not SOTA, is it reasonable?

prstrive commented 2 years ago

Maybe you can try the dynamic consistency checking strategy mentioned in EPP-MVSNet.