Research map / 3D vision / Depth and stereo

CLoSeR: Closing the Loop for Long-Context Streaming Reconstruction

NeurIPS 2026 · 3D vision › Depth and stereo · approach: Feed-forward 3D reconstruction · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Adds loop closure to the streaming feedforward reconstruction backbone LoGeR by retrieving loop candidates with SALAD descriptors, building loop-conditioned windows for relative poses, and optimizing all poses on SE(3) with a Huber-robust pose graph, reporting large ATE reductions on kilometer-scale benchmarks.

Problem

Long-context streaming feedforward 3D reconstruction from monocular video suffers tracking drift from error accumulation, especially when revisiting regions, and submap-based loop closure needs Sim(3)/SL(4) pose graphs because per-submap scales are inconsistent.

Approach

Exploit LoGeR's globally consistent scale and its non-contiguous window inference: append loop-conditioned windows (current plus revisited frames) to get relative poses directly in the global frame, then run SE(3) pose graph optimization with sequential and Huber-robust loop constraints.

Key results

• On VBR, CLoSeR achieves 11.15 m average ATE RMSE versus 36.16 m for LoGeR and 31.75 m for LoGeR*.
• On KITTI Odometry, CLoSeR reaches 12.79 m average ATE RMSE versus 25.44 m for LoGeR and 18.65 m for LoGeR*.
• On Oxford Spires, CLoSeR averages 5.26 m ATE RMSE versus 6.88 m for LoGeR, with runtime overhead of about 6% in FPS (7.12 vs 7.60).

Surprising findings

• Sim(3) and SL(4) pose graph variants performed worse than the SE(3) formulation, as extra degrees of freedom introduced ambiguity.
• The paper's runtime appendix reports 12.66 m ATE in text but 12.79 m in the table for CLoSeR on KITTI.

Limitations

• Fails when the feedforward frontend (LoGeR) completely breaks down, e.g., in highly dynamic environments or under aggressive camera motion.
• Does not explicitly handle dynamic objects, and is not competitive with dynamics-aware methods on DROID-W in the best case.
• Relies on LoGeR's globally consistent scale; the SE(3) formulation may not hold for backbones without this property.

Experimental setup

Datasets: VBR KITTI Odometry Oxford Spires DROID-W

Benchmarks: VBR KITTI Odometry DROID-W Oxford Spires

Metrics: ATE RMSE FPS peak GPU memory relative scale error

Baselines: DROID-SLAM DPV-SLAM DPV-SLAM++ VGGT-Long VGGT-SLAM VGGT-SLAM 2.0 CUT3R TTT3R LingBot-Map LoGeR LoGeR* InfiniteVGGT

Builds on: LoGeR VGGT SALAD DINO Huber loss

Closest papers by meaning

• Rethinking the State Update Gate for Long-Sequence Recurrent 3D Reconstruction (NeurIPS 2026 · similarity 0.88)
• IGGT4D: Streaming 4D Instance-Grounded Geometry Transformer (NeurIPS 2026 · similarity 0.86)
• FrameVGGT: Coherence-Preserving Memory for Bounded Streaming Geometry (NeurIPS 2026 · similarity 0.85)
• GHOST: Geometry-Hierarchical Online Streaming Token Eviction for Efficient 3D Reconstruction (NeurIPS 2026 · similarity 0.85)
• Restage4D: Reanimating Deformable 3D Reconstruction from a Single Video (NeurIPS 2025 · similarity 0.85)
• Fin3R: Fine-tuning Feed-forward 3D Reconstruction Models via Monocular Knowledge Distillation (NeurIPS 2025 · similarity 0.84)