The Gaussian reconstructed using gsplat on my own dataset was of poor quality, with a lot of noise and floating objects, resulting in poor subsequent semantic segmentation.

[cz2323]

How did you manage to reconstruct the Gaussian constellations to such high quality? My Gaussian constellations reconstructed using GSplat on my own dataset weren't very good; they had a lot of noise and floating objects, leading to poor subsequent semantic segmentation.

[cz2323]

I used photos taken at specific locations with my phone, and the image quality is fine. How can I optimize the parameters used in the GSplat reconstruction?

[unique1i]

We tried to obtain better quality 3DGS of both rendering and scene geometry at the start. For the scenes, we initialize the 3DGS from the provided geometry, for example the mesh provided in the ScanNet++ dataset, and apply the depth loss during optimization.

I attached the command I used before for running Gsplat. Here i edited the gsplat repo with a few more args, the edited repo is released at https://github.com/GaussianWorld/SceneSplat_Benchmark/tree/main/3dgs_optimization. If it is a custom scene without captured depth/geometry, the optimized 3DGS quality might not be ideal. If so please try more recent 3DGS optimization methods or using learned depth priors like DepthAnything3 for geometry supervision.

# Scannet++ MCMC
python -u examples/simple_trainer.py \
    mcmc \
    --strategy.cap-max 1500000 \
    --data_dir "$SCENE_PATH" \
    --result_dir "$OUTPUT_DIR" \
    --dataset_name "$DATASET_NAME" \
    --depth_loss \
    --depth_lambda 1 \
    --opacity_reg 0.0 \
    --scale_reg 0.02 \
    --max_steps 30000 \
    --disable_viewer \
    --prune_by_bbox

[cz2323]

Thank you for your reply, I will try the repo you provided.