Research map / 3D vision / Depth and stereo
SuperPrimitive: Scene Reconstruction at a Primitive Level
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.
Introduces SuperPrimitives, SAM-based image segments with normal-integrated unscaled depth whose per-segment scales are optimized via photometric multi-view alignment, applied to zero-shot depth completion, few-view SfM, and monocular VO.
Problem
Joint camera pose and dense geometry estimation from images or monocular video is hard due to computational cost and visual ambiguities; pixel-level multi-view methods fail on textureless or specular surfaces.
Approach
Represent an image as a set of local segments (SAM masks) each carrying a dense unscaled depth obtained by integrating single-view surface normals, then optimise only one depth scale per segment plus camera pose using photometric consistency, instead of per-pixel depth.
Key results
- Zero-shot depth completion on VOID (150 points): iRMSE 83.40 vs VI-Depth (DPT-Hybrid+NYUv2) 85.20, outperforming zero-shot group on three of four metrics.
- Monocular VO on TUM Freiburg 1 (8 sequences): mean ATE 0.153, best among compared methods (DepthCov 0.167, TartanVO 0.188).
- Few-view SfM on ScanNet: depth quality saturates after about 2 supporting views and consistently beats DeepV2D.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SuperPrimitives for zero-shot depth completion instead of VI-Depth | Depth completion · VOID 150 points | MAE: 104.96 -> 109.0 | Hurt |
| Use SuperPrimitives for zero-shot depth completion instead of VI-Depth | Depth completion · VOID 150 points | RMSE: 253.14 -> 204.15 | Helped |
| Use SuperPrimitives for zero-shot depth completion instead of VI-Depth | Depth completion · VOID 150 points | iMAE: 55.90 -> 47.32 | Helped |
| Use SuperPrimitives for zero-shot depth completion instead of VI-Depth | Depth completion · VOID 150 points | iRMSE: 85.20 -> 83.40 | Helped |
| Estimate depth with SuperPrimitives rather than DeepV2D | Few-view SfM · ScanNet test | iMAE | Helped |
| Estimate depth with SuperPrimitives rather than DeepV2D | Few-view SfM · ScanNet test | iRMSE | Helped |
| Use the SuperPrimitive monocular odometry system instead of DepthCov | Monocular visual odometry · TUM Freiburg 1 mean | ATE: 0.167 -> 0.153 | Helped |
Surprising findings
- Depth completion reaches performance comparable to a method trained on the task without any depth-completion training, using only zero-shot priors.
Limitations
- Monocular VO tracking is a proof-of-concept at 2-3 FPS without global bundle adjustment.
- Assumes geometry is continuous within each SAM segment; incorrect segments cannot be compensated after integration.
- No explicit occlusion handling in photometric alignment.
Experimental setup
Datasets: VOID ScanNet TUM RGB-D HyperSim
Benchmarks: VOID ScanNet TUM RGB-D
Metrics: MAE RMSE iMAE iRMSE ATE
Baselines: VI-Depth DeepV2D TartanVO DeepFactors DepthCov KBNet MonDi ScaffNet NLSPN VOICED
Builds on: Segment Anything DTAM DSO DeepV2D VI-Depth
Closest papers by meaning
- No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images (ICLR 2025 · similarity 0.87)
- SuperNormal: Neural Surface Reconstruction via Multi-View Normal Integration (CVPR 2024 · similarity 0.86)
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.86)
- From Correspondences to Pose: Non-minimal Certifiably Optimal Relative Pose without Disambiguation (CVPR 2024 · similarity 0.85)
- ODHSR: Online Dense 3D Reconstruction of Humans and Scenes from Monocular Videos (CVPR 2025 · similarity 0.85)
- Sparse-view Pose Estimation and Reconstruction via Analysis by Generative Synthesis (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.