Research map / Video / Video generation
Unified Video Dense Prediction from Disjoint Data
Yihong Sun, Seoung Wug Oh, Jiahui Huang, Bharath Hariharan, Joon-Young Lee
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 |
|---|---|---|---|
| use diffusion-prior backbone instead of frozen DINOv3-H with DPT heads | unified dense predictor · IIW albedo (OOD) | WHDR 15%: 0.289 -> 0.207 | Improved |
| use diffusion-prior backbone instead of frozen DINOv3-H with DPT heads | unified dense predictor · SAW shading (OOD) | P@0.7: 87.8 -> 94.6 | Improved |
| use diffusion-prior backbone instead of frozen DINOv3-H with DPT heads | unified dense predictor · Mapillary instance boundary (OOD) | odsF: 55.0 -> 57.2 | Improved |
| use diffusion-prior backbone instead of frozen DINOv3-H with DPT heads | unified dense predictor · NYUv2 depth | AbsRel: 0.092 -> 0.059 | Improved |
| use specialist diffusion model instead of DINOv3-H under color jittering | specialist · Hypersim val color jitter, depth | AbsRel deviation: 0.034 -> 0.011 | Improved |
| use specialist diffusion model instead of DINOv3-H under color jittering | specialist · Hypersim val color jitter, normals | mean error deviation: 5.8 -> 2.2 | Improved |
| distill all eight specialists into one unified backbone (cross-task regularization from depth) | surface normal specialist · ScanNet surface normals | mean angular error: 15.4 -> 14.6 | Improved |
| add projector-only fine-tuning after latent l1 distillation | UniD unified · Cityscapes semantic seg | mIoU: 37.8 -> 58.4 | Improved |
| add projector-only fine-tuning after latent l1 distillation | UniD unified · ADE20K semantic seg (OOD) | mIoU: 35.7 -> 44.8 | Improved |
| add video training (all videos as independent images otherwise), no inference memory (ablation) | UniD unified · ScanNet video depth (temporal) | AbsRel: 0.121 -> 0.109 | Improved |
| add video training, no inference memory (ablation) | UniD unified · VIPSeg semantic temporal consistency | mVC-4: 85.9 -> 86.5 | Improved |
| add extended self-attention inference memory [16,1] without video training (ablation) | UniD unified · VIPSeg semantic temporal consistency | mVC-4: 85.9 -> 89.1 | Improved |
Closest papers by meaning
- Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data (CVPR 2024 · similarity 0.88)
- PARTS: Unsupervised Segmentation With Slots, Attention and Independence Maximization (ICCV 2021 · similarity 0.87)
- A Simple yet Universal Framework for Depth Completion (NeurIPS 2024 · similarity 0.87)
- Video Depth Anything: Consistent Depth Estimation for Super-Long Videos (CVPR 2025 · similarity 0.87)
- Unified Dense Prediction of Video Diffusion (CVPR 2025 · similarity 0.86)
- Learning Temporally Consistent Video Depth from Video Diffusion Priors (CVPR 2025 · similarity 0.86)
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.
