[Question] Guidance on reproducing ScanNet results & discrepancies in data loading logic

[wdseh]

Hi authors,

Thank you for this impressive work! I am currently trying to reproduce the ScanNet results reported in the paper using your provided codebase.

While the lerf_ovs pipeline works well, I noticed that the current codebase seems to lack a specific entry point or configuration for ScanNet. I have checked the scene/dataset_readers.py and found that the readCamerasFromTransforms function (used for JSON/Blender format) seems to expect feature files ending in .pt (e.g., _fmap_CxHxW.pt), whereas the OpenGaussian dataset typically provides .npy files.

Could you please clarify the following points regarding the ScanNet pipeline:

Pipeline Consistency: Does the ScanNet experiment follow the exact same training pipeline as LeRF (i.e., Stage 1 Instance Field -> Stage 2 Mapping)? Or are there specific modifications required for ScanNet data?

Data Loading: Should we modify the readCamerasFromTransforms loader to support OpenGaussian's .npy format (similar to how readColmapCameras handles it), or is there a separate branch/script for ScanNet?

Hyperparameters: Could you share the training hyperparameters used for ScanNet (e.g., iterations, resolution downsampling, SAM levels)?

It would be greatly appreciated if you could release the training scripts for ScanNet or provide some guidance on how to adapt the current dataloader.

Thanks for your help!

[Runsong123]

Hi authors,

Thank you for this impressive work! I am currently trying to reproduce the ScanNet results reported in the paper using your provided codebase.

While the lerf_ovs pipeline works well, I noticed that the current codebase seems to lack a specific entry point or configuration for ScanNet. I have checked the scene/dataset_readers.py and found that the readCamerasFromTransforms function (used for JSON/Blender format) seems to expect feature files ending in .pt (e.g., _fmap_CxHxW.pt), whereas the OpenGaussian dataset typically provides .npy files.

Could you please clarify the following points regarding the ScanNet pipeline:

Pipeline Consistency: Does the ScanNet experiment follow the exact same training pipeline as LeRF (i.e., Stage 1 Instance Field -> Stage 2 Mapping)? Or are there specific modifications required for ScanNet data?

Data Loading: Should we modify the readCamerasFromTransforms loader to support OpenGaussian's .npy format (similar to how readColmapCameras handles it), or is there a separate branch/script for ScanNet?

Hyperparameters: Could you share the training hyperparameters used for ScanNet (e.g., iterations, resolution downsampling, SAM levels)?

It would be greatly appreciated if you could release the training scripts for ScanNet or provide some guidance on how to adapt the current dataloader.

Thanks for your help!

Hi, thanks for your interest, and sorry for the late reply. This current repository supports direct training, inference, and evaluation on the LeRF dataset. Regarding your questions on the ScanNet dataset: 1) The training pipeline still follows the same two-stage process. 2) For data loading, it keeps the same as in OpenGaussian, and you may need to first download the preprocessed dataset from OpenGaussian. 3) For hyperparameters, we use 30K iterations, a resolution downsampling value of 2, and a SAM level of 0 by default.

Hope this helps.