tsdf_mesh_tnt.py killed

[Austin-Chi]

Hi, when I try to reproduce your result on TnT dataset this happened:

PYTHONPATH=./ python mesh_extract/tsdf_mesh_tnt.py outp
ut/tnt/Barn --num_cluster 1 --use_depth_filter --sdf_trunc_scale 4.0
Rendering output/tnt/Barn
Read dataset in COLMAP format.
images_folder=/mnt/HDD_1/datasets/TNT_GOF/TrainingSet/Barn/images
res_downscale=2.0
res_width=0
computing nearest_id
Rendering progress: 100%|████████████████████████████████████████████| 410/410 [00:02<00:00, 161.05it/s]
TSDF Fusion progress: 20%|████████▊ | 82/410 [03:50<26:57, 4.93s/it]Killed

I am not sure if it is because of CPU or GPU RAM issue (currently I have a RTX5090 with 32GB RAM so maybe not this issue). Can you provide what resources are needed to reproduce your experiments? Thanks.

Another question is: You mentioned in READMe that the Courthouse we should use the estimation from 2DGS, do you mean just replace the Courthouse_COLMAP_SfM.log with the one from 2DGS?

[Fictionarry]

Hi, I think CPU RAM should be the problem. TSDF could have a high peak CPU mem during the process. I conduct the experiments on the server with 128GB RAM, and the current mesh extraction script sometimes costs over 30% or 50% RAM in some cases. But the peak cost should not last long. So adding some swap may work in your case. Offloading the ground truth images can also help I guess. These are actually not necessary during the extraction but cost much.

Just to replace the whole scene folder of courthouse with the one from 2DGS's. Its estimated poses for this scene are slightly more accurate than processed by the colmap script. But there are some alignment errors in some other scenes provided by 2DGS, so I still suggest processing the TnT dataset following Neuralangelo for the remaining.

[Fictionarry]

Hi, I think CPU RAM should be the problem. TSDF could have a high peak CPU mem during the process. I conduct the experiments on the server with 128GB RAM, and the current mesh extraction script sometimes costs over 30% or 50% RAM in some cases. But the peak cost should not last long. So adding some swap may work in your case. Offloading the ground truth images can also help I guess. These are actually not necessary during the extraction but cost much.

Just to replace the whole scene folder of courthouse with the one from 2DGS's. Its estimated poses for this scene are slightly more accurate than processed by the colmap script. But there are some alignment errors in some other scenes provided by 2DGS, so I still suggest processing the TnT dataset following Neuralangelo for the remaining.

Oh just 30% for TnT, not 50%. Sorry for my misleading. The scene Barn cost up to 20% of 128GB cpu mem in my test.