Research map / 3D vision / Depth and stereo

CM-EVS: Sparse Panoramic RGB-D-Pose Data for Complete Scene Coverage

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.

Proposes COVER, a training-free conflict-aware greedy ERP viewpoint curator, and uses it to build CM-EVS, a sparse provenance-tracked panoramic RGB-D-pose dataset from 3D indoor assets plus re-encoded outdoor panoramas.

Problem

Panoramic RGB-D-pose training data built from 3D assets lacks a sparse, comparable, geometry-consistent sampling policy; dense trajectories are redundant, heuristics miss regions or produce depth-inconsistent views, and view-selection choices are rarely released.

Approach

Greedy fixed-budget ERP view selection that warps accumulated range-depth geometry into candidate probes, scores incremental coverage minus a depth-conflict penalty (s = G - λL), and releases per-step logs; claims a bounded-error greedy coverage guarantee under noisy proxy oracle.

Key results

• On scene_indoor_0012 at K=4, Greedy coverage reaches 0.1055 coverage vs 0.0096 for random-seeded, while CM-EVS (lambda=0.35) reaches 0.1032 coverage with conflict 0.0170 vs 0.0193 for greedy coverage.
• CM-EVS dataset has 36,373 curator-produced indoor ERP frames from 1,275 scenes with median ~25 frames per indoor scene, covering all 13 room-type buckets.
• Lambda sweep at K=30 on 10 Blender scenes: lambda=0.35 gives coverage 0.4127 and conflict 0.0175, versus lambda=0 giving coverage 0.1878 and conflict 0.1561.

Surprising findings

• Lambda=0 (pure coverage gain) yields lower coverage (0.1878) than lambda=0.2 (0.4322), collapsing to a high-conflict tight cluster.
• The paper reports the downstream task evaluation is deferred; results only concern curator coverage and conflict statistics, not model accuracy.

Limitations

• Evaluation covers curator coverage and conflict statistics only, with no downstream task accuracy.
• HM3D and ScanNet++ frames are not redistributed and must be regenerated locally under upstream terms.
• Outdoor sources are re-encoded full trajectories without COVER selection, so they lack provenance logs.

Experimental setup

Datasets: Blender indoor HM3D ScanNet++ TartanGround OB3D

Benchmarks: Blender indoor pilot scene_indoor_0012 10-scene Blender indoor pool cross-source Blender/HM3D/ScanNet++ pools

Metrics: coverage coverage per view conflict ratio room-type entropy Gini concentration

Baselines: Random-seeded Single-view probe Greedy coverage Low-conflict only Matterport3D Hypersim 360DVD Matrix-Pano

Builds on: greedy submodular maximization Max-k-Cover BlenderProc Habitat-Sim Datasheets for Datasets

Closest papers by meaning

• LEAP: Liberate Sparse-View 3D Modeling from Camera Poses (ICLR 2024 · similarity 0.87)
• SparseGNV: Generating Novel Views of Indoor Scenes with Sparse RGB-D Images (AAAI 2024 · similarity 0.87)
• Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.86)
• PanoWorld: Geometry-Consistent Panoramic Video World Modeling (NeurIPS 2026 · similarity 0.86)
• MvFFN: Multi-view Floor-Plan Feed-Forward Network for Unposed Wide-Baseline Panorama Layout Reconstruction (NeurIPS 2026 · similarity 0.86)
• Splatter-360: Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images (CVPR 2025 · similarity 0.86)