JeffWang987 / MVSTER

[ECCV 2022] MVSTER: Epipolar Transformer for Efficient Multi-View Stereo
MIT License
189 stars 13 forks source link

Are DTU results reported in the paper trained on 640 or 1280? #2

Closed ewrfcas closed 2 years ago

ewrfcas commented 2 years ago

Thanks for your good work! I have not seen any discussion of the input resolution in the paper.

MaybeOjbk commented 2 years ago

I'm also curious about the Ablation Study of your paper? Are the depths trained in 1600 x1200 or 640 x 512? And also I want to know that does the high-resolution training strategy boost the depth quality rapidly

cnexah commented 2 years ago

Same question: is training on 1600 x1200 necessary? or we just train on 640 x 512 and test on 1600 x1200?

JeffWang987 commented 2 years ago

To boost the performance, MVSTER* is trained on 1600 x1200 and then tested on 1600 x1200. In the ablation, we train on 640 x 512 and test on 864 x 1152.

cnexah commented 2 years ago

Thank you!