the results in Tanks and Temples - Githubissues

prstrive / UniMVSNet

[CVPR 2022] Rethinking Depth Estimation for Multi-View Stereo: A Unified Representation

MIT License

231 stars 12 forks source link

the results in Tanks and Temples #7

Closed xingyi-li closed 2 years ago

xingyi-li commented 2 years ago

Thanks for your generous open source. We apply unimvsnet_blendedmvs.ckpt to reproduce the results of Tanks and Temples via scripts/tank_test.sh. Results are approaching in advanced while those of intermediate drop considerably. Is there any different setting between intermediate and advanced，e.g., inverse depth, num_view, ndepths, interval_ratio ?

The full results are as follow:

F-score:

Auditorium: 27.28 Ballroom: 44.05 Courtroom: 39.13 Family: 34.35 Francis: 24.06 Horse: 06.13 Lighthouse: 58.74 M60: 57.29 Museum: 52.20 Palace: 32.27 Panther: 57.88 Playground: 57.57 Temple: 33.45 Train: 45.85

Mean intermediate: 42.73 Mean advanced: 38.06

prstrive commented 2 years ago

I guess you may not use the short depth range version of the camera parameters in Intermediate set as we described here.