Make DA3 W1 the uncalibrated calibration default

[Zador-Pataki]

Motivation

The uncalibrated pipeline uses batched GeoCalib initialization and raw VGC. Reuse Depth Anything 3 (DA3) W1 depth inference to predict intrinsics and regularize focal lengths through VGC and bundle adjustment.

Changes

• Make DA3 W1 the uncalibrated default; initialize focal lengths from the median prediction and retain original per-image predictions as prior targets.
• Whiten log-focal residuals by the assumed log-focal standard deviation, then apply Cauchy loss in VGC, Huber in normal BA, and Cauchy during annealing.
• Keep principal points fixed at image center and GeoCalib available as an alternative.

Performance

Accuracy remains comparable on supported benchmarks, with substantial improvements observed on an external benchmark.

Reusing DA3 depth inference for focal estimates also eliminates the separate GeoCalib pass, reducing frontend inference cost.

Changes to the README benchmark tables

Differences may also reflect run-to-run noise and other pipeline changes since the initial release.

WATE-AUC	10m / 0.5m	25m / 1.25m	50m / 2.5m	100m / 5m	Full / 5%
LaMAR	92.10 → 92.16	92.75 → 92.62	91.29 → 91.29	90.81 → 90.67	90.22 → 90.20
CroCoDL	94.47 → 94.90	95.90 → 96.18	96.23 → 96.44	94.87 → 94.21	96.11 → 96.60

Pose AUC	5cm	10cm	50cm	1m	10m
EuRoC	11.39 → 11.96	31.40 → 32.05	80.50 → 81.67	90.25 → 90.84	99.03 → 99.08
ETH3D-SLAM	49.16 → 50.15	65.98 → 66.99	88.49 → 88.32	92.98 → 92.76	98.59 → 98.83