Research map / Video / Video-language understanding
Beyond Simple Edits: Composed Video Retrieval with Dense Modifications
Omkar Thawakar, Dmitry Demidov, Ritesh Thawkar, Rao Muhammad Anwer, Mubarak Shah, Fahad Shahbaz Khan, Salman Khan
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 dense descriptions and dense modification text in training (ablation) | Ours · Dense-WebVid-CoVR test | R@1: 60.4 -> 71.2 | Improved |
| add dense modification text (descriptions kept WebVid-CoVR) (ablation) | Ours · Dense-WebVid-CoVR test | R@1: 60.4 -> 61.2 | Improved |
| add dense descriptions (modification text kept WebVid-CoVR) (ablation) | Ours · Dense-WebVid-CoVR test | R@1: 60.4 -> 63.8 | Improved |
| use dense modification text at inference (ablation) | Ours · Dense-WebVid-CoVR test | R@1: 68.88 -> 71.26 | Improved |
| use dense video descriptions at inference (ablation) | Ours · Dense-WebVid-CoVR test | R@1: 66.08 -> 71.26 | Improved |
| use unified weighted-mean fusion instead of addition (ablation) | Ours · Dense-WebVid-CoVR test | R@1: 69.72 -> 71.26 | Improved |
| use unified weighted-mean fusion instead of cross-attention for video and description (ablation) | Ours · Dense-WebVid-CoVR test | R@1: 70.13 -> 71.26 | Improved |
| replace pairwise fusion with unified single grounding encoder fusion (no dense training) | Thawakar et al. · Dense-WebVid-CoVR test (CA, no dense training) | R@1: 39.20 -> 48.08 | Improved |
| replace pairwise fusion with unified fusion (trained on dense data) | Thawakar et al. · Dense-WebVid-CoVR test | R@1: 67.86 -> 71.26 | Improved |
| train on Dense-WebVid-CoVR rather than not | Ours · Dense-WebVid-CoVR test (Text only) | R@1: 24.12 -> 50.12 | Improved |
| unified fusion vs pairwise fusion inference speed | Thawakar et al. · Dense-WebVid-CoVR | inference speed: 3x faster | Improved |
| apply unified fusion to composed image retrieval, training on CIRR | Thawakar et al. · CIRR test | R@1: 51.03 -> 56.30 | Improved |
Closest papers by meaning
- Composed Video Retrieval via Enriched Context and Discriminative Embeddings (CVPR 2024 · similarity 0.91)
- CoVR: Learning Composed Video Retrieval from Web Video Captions (AAAI 2024 · similarity 0.91)
- From Play to Replay: Composed Video Retrieval for Temporally Fine-Grained Videos (NeurIPS 2025 · similarity 0.91)
- Mitigating Semantic Collapse in Partially Relevant Video Retrieval (NeurIPS 2025 · similarity 0.89)
- Video-RAG: Visually-aligned Retrieval-Augmented Long Video Comprehension (NeurIPS 2025 · similarity 0.88)
- Towards Efficient and Effective Text-to-Video Retrieval with Coarse-to-Fine Visual Representation Learning (AAAI 2024 · similarity 0.88)
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.
