[Zador-Pataki]

Motivation

VidMap maintains its own GP and BA depth residuals. The proposed COLMAP depth appenders and cost functors provide these operations with log-scale inputs, allowing VidMap to remove the duplicated residual implementations.

Changes

• Group BA observations by shared loss and call the upstream appender from Python, preserving the existing log-scale prior and variance-floor weights.
• Use the upstream GP appender for ordinary observations. Add a small native adapter for existing loop-closure associations and extra-center observations, using upstream cost functors.
• Store, optimize, and regularize log depth scales in both GP and BA.
• Use linear depth residuals in GP1 and upstream log-linear residuals in GP2, removing the superseded depth options.
• Pin the proposed COLMAP depth API to log-scale-only commit 6cde3c7f.

GP's log-scale regularization and upstream transition and uncertainty handling intentionally change its objective. BA retains its intended objective. Existing warmup and center initialization are unchanged.

Dependency

Pending dependency: this draft pins an unpublished COLMAP depth implementation (6cde3c7f). Public builds are blocked until that implementation is published and the pin is updated. Validation against the final dependency revision is also pending.