Skip to content

Latest commit

 

History

History
63 lines (51 loc) · 2.05 KB

File metadata and controls

63 lines (51 loc) · 2.05 KB

Geometry Priors (MoGe)

Generate dense depth and normal priors using MoGe to improve FriendlySplat's geometry reconstruction.

Installation

pip install -e ".[priors]" --no-build-isolation

Quick Start

Run the installed CLI entrypoint. This is the recommended standard command:

fs-prior-moge \
  --data-dir /path/to/scene \
  --factor 1 \
  --model-id Ruicheng/moge-2-vitl-normal \
  --out-normal-dir moge_normal \
  --out-depth-dir moge_depth \
  --align-depth-with-colmap \
  --save-depth \
  --verbose

Key Parameters Explained

  • --align-depth-with-colmap: enabled by default and crucial for usable depth. MoGe predicts relative depth. This flag fits it to your COLMAP sparse scene under sparse/0/, turning it into scene-consistent depth.
  • --save-depth: by default the script always exports normals, but depth .npy files are only written when you pass this flag.
  • --save-sky-mask: optionally export invalid-depth / sky-like masks under sky_mask/, which can be consumed by FriendlySplat via data.sky_mask_dir_name.
  • --factor: controls which image folder is used, such as images_2/ or images_2p5/. It must match your training --data.data-factor, otherwise FriendlySplat will hit image-prior shape mismatches.
  • --model-id: HuggingFace model weights. Ruicheng/moge-2-vitl-normal is the default and recommended high-quality model.
  • --remove-depth-edge: enabled by default. It masks sharp depth discontinuities to reduce floating artifacts during splat training.
  • --verbose: prints per-image alignment details and is useful when depth fitting behaves unexpectedly.

Use in Training

Once generated, pass the output directory names to training:

fs-train --io.data-dir /path/to/scene \
  --data.data-factor 1 \
  --data.normal-dir-name moge_normal \
  --data.depth-dir-name moge_depth \
  --data.sky-mask-dir-name sky_mask

Tip: if you only need normals, skip --save-depth and add --no-align-depth-with-colmap to avoid the heavier COLMAP depth fitting step.