Multi-Reference Geometric Risk and Sparse Depth Refinement for Streaming 3D Reconstruction

Streaming feed-forward 3D reconstruction produces depth predictions at successive frames, but cross-view geometric consistency and local depth accuracy may vary. We present a training-free, output-level approach that combines Multi-Reference Geometric Risk assessment with global-referenced conservative sparse depth refinement, leaving the pretrained backbone, its predicted poses, and its streaming state unchanged. Risk derived from predicted depth, intrinsics, and poses achieves a median Spearman correlation of 0.635 with scale-aligned depth error across 11 public scenes and a correlation of 0.556 on the real-robot Sequence E. With matched 2% sparse depth anchors, local refinement reduces Sequence E held-out AbsRel by 40.69% relative to Sparse Global. Spatial refinement also improves mean held-out depth error in all 55 public scene–policy comparisons and all 200 predefined fixed-mask scene–policy–seed comparisons. Retrospective risk-based frame ranking provides diagnostic benefits under reference-scale normalization, but the observed allocation gains are protocol-dependent and do not establish consistent improvements under historical sparse-scale propagation or deployable online scheduling. In an independent Bonn Leica comparison using shared ground-truth camera poses, Spatial increases recall at a 5 cm threshold but has lower precision and F-score than Sparse Global. These findings establish the value of conservative sparse depth refinement and multi-reference risk as complementary geometric-quality evidence, while distinguishing depth accuracy from downstream 3D geometric performance.