search

TruongKhang / cds-mvsnet

[ICLR2022] Curvature-guided dynamic scale networks for Multi-view Stereo
118 stars 6 forks source link

Evaluate the results #11

Closed Sun-Xinnnnn closed 2 years ago

Sun-Xinnnnn commented 2 years ago

Hello, author. The result of reconstruction of DTU model in prepareed provided by you is not satisfactory. I also set prob_threshold. What is the reason?

python test.py --dataset dtu --batch_size 1 --testpath $TESTPATH --testlist $TESTLIST --resume $CKPT_FILE --outdir $OUTDIR --interval_scale 1.06 --num_view 5 --numdepth 192 --max_h 1152 --max_w 1536 --filter_method gipuma --disp_threshold 0.1 --num_consistent 2 --prob_threshold 0.19 git

TruongKhang commented 2 years ago

@zhang-snowy, which pre-trained model did you use? You can remove prob_threshold, it does not affect much for DTU evaluation. Also, I uploaded the 3D reconstructed models for DTU in README. If you find your reconstructed models similarly, then they would be fine! Your image is scene 77, in my experience, most methods cannot reconstruct accurately for this scene because of large untextured regions.

Sun-Xinnnnn commented 2 years ago

@TruongKhang Thank you very much for your reply, I use your CDS - mvsnet/tree/main/pretrained under the training model of fusion

TruongKhang commented 2 years ago

@zhang-snowy, then your produced result is reasonable. As I explained above, the scene 77 is hard to reconstruct completely because of large untextured scenes. Actually, if you run the Matlab evaluation code, this scene also gave a lowest evaluated metrics. And just one small thing, if you like our repo, please give me 1 star :D

Sun-Xinnnnn commented 2 years ago

OK ! ! ! Thank you very much for your reply ! ! ! : D