Incomplete mesh reconstruction result

[Nole21]

Hello, jiahe. Thanks for sharing such awesome work! I observed that the released huggingface results on TNT is different from the visualization in the paper. The reconstructed mesh is incomplete which is not observed in other surface reconstruction methods based on gaussian splatting. For example, the roof of the barn is a visible part of the scene, but the mesh is incomplete, especially on the Caterpillar scene. Look forward to your reply.
Thanks, Good luck.

[Fictionarry]

Hi, thanks for your interest! These meshes are truncated strictly in TSDF to serve the evaluation, where only the precisely accurate points remain. You can just place the flag --sdf_trunc_scale 8.0 during running mesh_extract/tsdf_mesh_tnt.py to reproduce the complete meshes shown at huggingface here https://huggingface.co/Fictionary/GeoSVR/tree/main/meshes_complete/TnT. Results of DTU can be referred to this issue #10.

[Nole21]

Thanks for your reply. Do you mean the metrics reported on the paper are based on incomplete mesh(without the flag --sdf_trunc_scale 8.0)?

[Fictionarry]

Thanks for your reply. Do you mean the metrics reported on the paper are based on incomplete mesh(without the flag --sdf_trunc_scale 8.0)?

Yes, both the F1 score and Chamfer distance are the balance between precision and completeness, so the absolutely complete mesh may not lead to the best metric. In my observation, it's shown a traditional practice for the surface reconstruction methods to only measure the points that are confidently accurate. Another problem is that the ground truths are also not pretty complete. The reconstructed points not provided in the ground truth can partially be considered incorrect (in DTU) or harm the point cloud alignment (in TnT). The modified TnT evaluation script provided by 2DGS and GOF seems can relieve this problem, but I'm not fully sure about it.