Prior-Driven Enhancements in 3D Gaussian Splatting: Normals and Depths Regularization
ISPRS Geospatial Week 2025 路 Oral 路 Gyeonggwan Lee, Seunghwan Hong, Junghun Suh (Kakao Mobility)
Project page 路 Code 路 Paper
Trained 3D Gaussian Splatting models for the Tanks and Temples rows of Table 1 (Train, Horse), for both vanilla 3DGS and ours, with Gaussians initialized from each of the three SfM pipelines (COLMAP, SP-SG, LoFTR). The Parking lots and Street-view models are not released, because they are built from in-house imagery.
Files
<method>/<sfm>/<scene>/ holds point_cloud.ply (the 30k-iteration Gaussians), cameras.json, and
the cfg_args the model was trained with. MD5 sums are in checkpoints.md5.
Table 1 of the paper (PSNR / SSIM / LPIPS on the held-out test views, every 8th image):
| Scene | SfM | 3DGS | Ours |
|---|---|---|---|
| Train | COLMAP | 21.10 / 0.802 / 0.218 | 21.97 / 0.799 / 0.252 |
| Train | SP-SG | 21.10 / 0.750 / 0.282 | 21.32 / 0.749 / 0.287 |
| Train | LoFTR | 20.97 / 0.767 / 0.274 | 21.11 / 0.759 / 0.291 |
| Horse | COLMAP | 24.18 / 0.889 / 0.239 | 25.50 / 0.903 / 0.153 |
| Horse | SP-SG | 21.01 / 0.802 / 0.239 | 21.12 / 0.801 / 0.246 |
| Horse | LoFTR | 23.39 / 0.870 / 0.174 | 24.80 / 0.881 / 0.165 |
Usage
The models load with the standard 3DGS viewers and with the repository's render.py, given the matching
Tanks and Temples scene directory:
hf download gandan-lee/pdigs --local-dir models
python render.py -m models/ours/loftr/Horse -s $DATA/defree_sfm/Horse --skip_train
License
Derived from 3D Gaussian Splatting; released under the Gaussian-Splatting License (non-commercial research and evaluation use). The Tanks and Temples imagery the models are fitted to is subject to its own terms.
Citation
@article{lee2025prior,
title = {Prior-Driven Enhancements in {3D} {Gaussian} Splatting: Normals and Depths Regularization},
author = {Lee, Gyeonggwan and Hong, Seunghwan and Suh, Junghun},
journal = {The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences},
volume = {XLVIII-G-2025},
pages = {891--897},
year = {2025},
doi = {10.5194/isprs-archives-XLVIII-G-2025-891-2025}
}