Reconstructing a 3D world from a long video is hard because camera position errors grow over time and misalign the map. Many systems limit drift by optimizing only a recent window, but still make mistakes.
MetricTrail is a streaming system: it processes frames as they arrive. It uses single-image metric depth, a per-frame distance estimate in real units, to seed and softly correct geometry. It also strengthens point matching with frozen pretrained visual features. The correspondence network is not retrained, and the optimization window stays small.
On three benchmarks, it achieved the lowest average camera path error among compared methods and produced coherent dense models from ordinary videos. This shows strong image-level priors can keep efficient local estimation accurate over long sequences, a practical benefit for robotics and large-scale mapping.