Research map / 3D vision / Depth and stereo
GS-ASM: 2DGS-Supervised Active Stereo Matching
Zhengling Wu, Rongfeng Lu, Quan Chen, Longjian Zeng, Ming Lu, Yaoqi Sun, Yahong Chen, Baofeng Ji, Chenggang Yan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with 2DGS proxy-label supervision plus novel views and hybrid regularization instead of ActiveZero hybrid-domain training | RAFT stereo network · Simulated test set (Blender) | EPE (px): 0.4715 -> 0.1279 | Improved |
| train with GS-ASM instead of ActiveZero baseline | PSMNet · Simulated test set (Blender) | EPE (px): 0.4300 -> 0.2613 | Improved |
| train with GS-ASM instead of ActiveZero baseline | StereoNet · Simulated test set (Blender) | EPE (px): 0.5551 -> 0.4124 | Improved |
| add 2DGS proxy labels for supervised training (ablation) | ActiveZero baseline on RAFT · Simulated test set (Blender) | EPE (px): 0.4715 -> 0.3482 | Improved |
| add novel-viewpoint data augmentation from 2DGS (ablation) | Baseline with proxy labels on RAFT · Simulated test set (Blender) | EPE (px): 0.3482 -> 0.2736 | Improved |
| add hybrid supervision regularization (adaptive loss weights) (ablation) | Proxy labels + novel viewpoints on RAFT · Simulated test set (Blender) | EPE (px): 0.2736 -> 0.1279 | Improved |
| add hybrid supervision regularization (ablation) | Proxy labels + novel viewpoints on RAFT · Simulated test set (Blender) | 5px accuracy: 0.9945 -> 0.9955 | Improved |
| fit 2DGS on IR images captured at higher emitter intensity instead of zero emitter power (ablation) | active 2DGS · 2DGS reconstruction of IR images | PSNR: 33.23 -> 31.75 | Hurt |
| use IR images at higher emitter intensity for binary pattern extraction (ablation) | binary IR pattern extraction · 2DGS reconstruction of IR images | LPIPS: 0.0762 -> 0.3936 | Hurt |
| use D435 depth as supervision instead of 2DGS proxy labels | RAFT · Simulated test set (Blender) | EPE (px): 0.1279 (ours) vs 0.5428 (D435) | Hurt |
| use ActiveZero hybrid-domain training in real scenes | ActiveZero · Real-world scenes (qualitative) | depth quality | Hurt |
Closest papers by meaning
- ActiveStereoNet: End-to-End Self-Supervised Learning for Active Stereo Systems (ECCV 2018 · similarity 0.89)
- Active-Passive SimStereo - Benchmarking the Cross-Generalization Capabilities of Deep Learning-based Stereo Methods (NeurIPS 2022 · similarity 0.88)
- Active Event-based Stereo Vision (CVPR 2025 · similarity 0.87)
- Consistency-aware Self-Training for Iterative-based Stereo Matching (CVPR 2025 · similarity 0.86)
- GS-2DGS: Geometrically Supervised 2DGS for Reflective Object Reconstruction (CVPR 2025 · similarity 0.85)
- IndoorGS: Geometric Cues Guided Gaussian Splatting for Indoor Scene Reconstruction (CVPR 2025 · 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.
