Research map / Video / Video-language understanding
Stitch-a-Demo: Creating Video Demonstrations from Multistep Descriptions
Chi Hsuan Wu, Kumar Ashutosh, Kristen Grauman
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 |
|---|---|---|---|
| add LLM-created weakly supervised mixed procedures to training (ablation) | Stitch-a-Demo · SaD-VD | R@5: 0.14 -> 0.56 (vs no augmentation) | Improved |
| use LLM-created procedures instead of temporally-sampled combinations (ablation) | Stitch-a-Demo · SaD-VD | median rank: 10 -> 3.5 | Improved |
| add visual continuity negatives to correctness negatives (ablation) | Stitch-a-Demo · SaD-VD | R@5: 0.41 -> 0.45 | Improved |
| add object state negatives to correctness + continuity negatives (ablation) | Stitch-a-Demo · SaD-VD | R@5: 0.45 -> 0.52 (R@5); MedR 9 -> 5 | Mixed |
| use only visual continuity negatives (ablation) | Stitch-a-Demo · SaD-VD | R@5: 0.03 (MedR 171) | Hurt |
| add correctness negatives to continuity + object state negatives (ablation) | Stitch-a-Demo · SaD-VD | R@5: 0.21 -> 0.52 | Improved |
| add visual continuity negatives to correctness + object state negatives (ablation) | Stitch-a-Demo · SaD-VD | R@5: 0.54 -> 0.52 (R@5); 0.15 -> 0.22 (R@1) | Mixed |
| replace heuristic scoring (InternVideo similarity) with learned procedure evaluator | Stitch-a-Demo · Cooking SaD-MC | R@50: 0.55 -> 0.84 | Improved |
| apply Stitch-a-Demo to woodworking and gardening | Stitch-a-Demo · Woodworking/Gardening SaD-MC | R@50: 0.36 vs CoVR 0.34 (woodworking) | Mixed |
| stitch clips from multiple videos instead of retrieving original single video demonstration | Stitch-a-Demo · Cooking human preference | preference ratio: 83:17 | Improved |
| use retrieval instead of image generation (ShowHowTo) | Stitch-a-Demo · Cooking human preference | win rate, total: 0.98 | Improved |
| compare VBench subject consistency at step transitions vs InternVideo | Stitch-a-Demo · object-state consistency | VBench subject consistency: 0.62 -> 0.67 | Improved |
Closest papers by meaning
- ShowHowTo: Generating Scene-Conditioned Step-by-Step Visual Instructions (CVPR 2025 · similarity 0.88)
- Aligning Step-by-Step Instructional Diagrams to Video Demonstrations (CVPR 2023 · similarity 0.87)
- StepFormer: Self-Supervised Step Discovery and Localization in Instructional Videos (CVPR 2023 · similarity 0.86)
- Step Differences in Instructional Video (CVPR 2024 · similarity 0.85)
- Detours for Navigating Instructional Videos (CVPR 2024 · similarity 0.84)
- Needle In A Video Haystack: A Scalable Synthetic Evaluator for Video MLLMs (ICLR 2025 · 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.
