COLMAP Config

Trimmed from colmap_default.yaml. Same YAML format as the original — copy the block below into a .yaml file to use it.

Removed on purpose: experiment block, all feature extraction/matching detail except the GPU switches, all input/output path blocks, Hydra settings. Paths must now be supplied by the caller (CLI override, or a wrapper config) — this file alone is not runnable.

# ── Pipeline flags ────────────────────────────────────────────────────────────
random_seed: 0

# ── Stage toggles (can also be driven by pipeline.yaml) ───────────────────────
run:
  colmap_feat_extract_match: true
  colmap_reconstruct: true

# ── Feature extraction / matching ─────────────────────────────────────────────
# GPU instead of CPU. All other extraction/matching options dropped -> COLMAP defaults.
feat_extract_match:
  extraction:
    device: cuda
  matching:
    use_gpu: true

# ── Reconstruction ────────────────────────────────────────────────────────────
# Priors are loaded as a STARTING POINT only, not frozen: COLMAP refines both poses
# and intrinsics during bundle adjustment and may move them away from the priors.
reconstruction:
  mode: automatic              # triangulate | automatic
  load_prior_poses: true       # automatic mode only: seed with known robot poses
  load_prior_intrinsics: true  # seed with known camera intrinsics
  refine_intrinsics: true      # keep true -> intrinsics stay free, not locked to prior
  mapping_params:
    min_focal_length_ratio: 0.1
    max_focal_length_ratio: 10.0
    max_extra_param: .inf
    ba_local_max_num_iterations: 30
    ba_local_max_refinements: 3
    ba_global_max_num_iterations: 75
    num_threads: 1

Notes for whoever runs this

mode: automatic is required for the pose priors. load_prior_poses is ignored in triangulate mode, per the original comment. Changing mode silently drops the pose priors.
"Don't freeze" is refine_intrinsics: true. Setting it to false pins intrinsics to the prior values, which is the opposite of what this config is for.
ba_use_gpu is deliberately absent from mapping_params — bundle adjustment is left to whatever the implementation defaults to. Only feature extraction and matching are pinned to GPU. Do not add a ba_use_gpu key back in.