about mapping performance

[jinxxo-j]

Hello,

I have a question regarding the SPLAT-LOAM implementation.

I run python3 run.py slam /configs/ncd/quad-easy-mapping-gt.yaml

&&

python3 run.py mesh <path/to/result/folder>

&&

python3 run.py eval_recon <reference_pointcloud_file> <estimate_mesh>

The evaluation produced the results shown in the attached screenshot. Unfortunately, the reconstruction metrics differ markedly from those reported in the paper and repository.

Could you suggest possible reasons for this discrepancy?
For example, are there additional configuration settings, post‑processing steps, or dataset‑specific parameters required to reproduce the published mapping performance?

Thank you in advance for your assistance.

[EmanueleGiacomini]

Hi!
The default configuration for meshing is tuned for speed rather than accuracy. Additional info concerning the post-processing steps can be found via --help over the mesh subcommand:

python3 run.py mesh <path/to/result/folder> --help

Generally speaking, the meshing accuracy is driven by two factors:

• Poisson depth: default to 10, but experiments were carried out with a depth of 12 (gets more accurate but slower)
• KF samples: Number of points to sample per keyframe

Additionally, we noticed that mapping with smaller local models (mapping.lmodel_threshold_[ngaussians/nkeyframes]) and higher densification ratios (mapping.densify_percentage) can lead to better results.

An additional note concerning the repeatability of the experimental results shown in the paper: The current implementation is still behaving differently from the implementation we used to generate the results, so expect some differences.

We will correct this discrepancy as soon as possible!

[jinxxo-j]

Thank you for your prompt response.
When will you make the correction? I'll be waiting.

[jinxxo-j]

Could you please provide the mesh output files generated by OpenVDB, VoxBlox, N3-Mapping, and PIN-SLAM for the Newer College dataset (quad-easy sequence) and the Oxford Spires dataset (keble-college02, bodleian-library-02, observatory-01 sequences), so that we can reproduce the mesh reconstruction quality comparison reported in the paper? We suspect that aggregating multiple mesh files and evaluating on cropped regions may be contributing to the observed reconstruction underperformance.

[EmanueleGiacomini]

Hi,
Sorry for the late reply, I have been busy with my phd defense + several post-doc tasks.
I'll try to upload the meshes asap