Research map / 3D vision / 3D generation
RaysUp: Ultra-light Universal Feature Upsampling via Geometry-Aware Ray Representation
Ding Yuchuan, Linfei Li, Lin Zhang, Ying Shen
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 |
|---|---|---|---|
| replace image-grid RoPE/no PE with RayPE (ablation) | RaysUp · avg mIoU over 4 VFMs | mIoU: 81.05 -> 82.17 (None -> RayPE); RoPE 81.92 | Improved |
| use RayPE instead of RoPE (ablation) | RaysUp · avg mIoU over 4 VFMs | mIoU: 81.92 -> 82.17 | Improved |
| use RayPE instead of SinRays (ablation) | RaysUp · avg mIoU over 4 VFMs | mIoU: 81.59 -> 82.17 | Improved |
| use decoupled-branch guidance encoder instead of single-branch (ablation) | RaysUp · avg mIoU over 4 VFMs | mIoU / params: 81.87 -> 82.17; 0.266M -> 0.14M | Improved |
| increase guidance dim from 256 to 512 or 768 (ablation) | RaysUp · avg mIoU over 4 VFMs | mIoU: 82.17 -> 82.20 (512), 82.19 (768); params 0.14M -> 0.40M/0.89M | No clear effect |
| reduce guidance dim from 256 to 128 (ablation) | RaysUp · avg mIoU over 4 VFMs | mIoU: 82.17 -> 82.02 | Hurt |
| increase guidance conv blocks from L=1 to L=3 (ablation) | RaysUp · avg mIoU over 4 VFMs | mIoU: 82.17 -> 82.44; params 0.14M -> 0.27M | Improved |
| use DA3 predicted pose instead of identity pose (ablation) | RaysUp · avg mIoU over 4 VFMs | mIoU: 82.17 -> 82.41 (DA3-Small), 82.44 (DA3-Base) | Improved |
| add local-crop training strategy (ablation) | RaysUp · Cityscapes segmentation | mIoU: 61.88 -> 63.04 | Improved |
| RaysUp vs LoftUp on open-vocabulary segmentation | RaysUp · COCO-Stuff / Pascal-VOC / ADE20K / Cityscapes open-vocab | mIoU: COCO-Stuff 27.54 (LoftUp) vs 27.11 | Hurt |
| RaysUp vs AnyUp across backbones | RaysUp · Pascal-VOC segmentation across DINOv2/DINOv3/SigLIP2/PE Spatial | mIoU: e.g. DINOv3 ViT-S 81.97 -> 82.60; SigLIP2 ViT-B 78.70 -> 79.37 | Improved |
| RaysUp vs AnyUp efficiency | RaysUp · 224x224 upsampling | FPS / params: 11 -> 55 FPS; 0.87M -> 0.14M params | Improved |
Closest papers by meaning
- FoundSurface: Feedforward Ray-Based 3D Scene Surface Reconstruction from Unposed Images (NeurIPS 2026 · similarity 0.88)
- LoftUp: Learning a Coordinate-Based Feature Upsampler for Vision Foundation Models (ICCV 2025 · similarity 0.87)
- RayNet: Learning Volumetric 3D Reconstruction With Ray Potentials (CVPR 2018 · similarity 0.87)
- FeatUp: A Model-Agnostic Framework for Features at Any Resolution (ICLR 2024 · similarity 0.86)
- CRAYM: Neural Field Optimization via Camera RAY Matching (NeurIPS 2024 · similarity 0.84)
- UniSHARP: Universal Sharp Monocular View Synthesis (NeurIPS 2026 · similarity 0.84)
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.
