search

autonomousvision / sdfstudio

A Unified Framework for Surface Reconstruction
Apache License 2.0
1.9k stars 179 forks source link

Poor results from training geo-volsdf and geo-unisurf #272

Open wujinhu1999 opened 6 months ago

wujinhu1999 commented 6 months ago

Hello authors, thank you for your great work! When I use the command 'ns-train geo-volsdf --pipeline.model.sdf-field.inside-outside False --vis wandb --experiment-name geo-volsdf-dtu65 sdfstudio-data --data data/dtu/scan65 --load-pairs True' and 'ns-train geo-unisurf --pipeline.model.sdf-field.inside-outside False --vis wandb --experiment-name geo-unisurf-scan24 sdfstudio-data --data data/dtu/scan24 --load-pairs True' is used to train geo-volsdf and geo-unisurf respectively, the output results are very bad. This is the output of geo-volsdf training 100k: depth_100000_5cb796e4dfdfb7c1c650 img_100000_4b032ba46420590b6288 normal_100000_500b1725dfe5b36dba95 This is the output of geo-unisurf training 50k: depth_50000_41ca62657722f53d8b39 img_50000_d0cd6c4d53f72b9329f4 normal_50000_a3d92c5f011a314000da Looking forward to your reply!

niujinshuchong commented 5 months ago

Hi, this is strange, could you reproduce our results with the same training command here: https://github.com/autonomousvision/sdfstudio/blob/master/docs/sdfstudio-methods.md

wujinhu1999 commented 5 months ago

Hi, this is strange, could you reproduce our results with the same training command here: https://github.com/autonomousvision/sdfstudio/blob/master/docs/sdfstudio-methods.md I used the instructions you gave me. 1) ns-train geo-volsdf --pipeline.model.sdf-field.inside-outside False --vis wandb --experiment-name geo-volsdf-dtu65 sdfstudio-data --data data/dtu/scan65 --load-pairs True 2) ns-train geo-unisurf --pipeline.model.sdf-field.inside-outside False --vis wandb --experiment-name geo-unisurf-scan24 sdfstudio-data --data data/dtu/scan24 --load-pairs True