Research map / Video / Video-language understanding
Prompt Switch: Efficient CLIP Adaptation for Text-Video Retrieval
Chaorui Deng, Qi Chen, Pengda Qin, Da Chen, Qi Wu
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 Prompt Switch prompt cube to CLIP image encoder (ablation) | CLIP ViT-B/32 · MSR-VTT 1K-A | R@1: 43.1 -> 44.8 | Improved |
| add Prompt Aggregation on top of Prompt Switch (ablation) | Prompt Switch · MSR-VTT 1K-A | R@1: 44.8 -> 45.4 | Improved |
| add auxiliary captioning loss (ablation) | Prompt Switch + Prompt Aggregation · MSR-VTT 1K-A | R@1: 45.4 -> 46.1 | Improved |
| use Prompt Aggregation instead of Temporal Transformer (ablation) | Prompt Switch · MSR-VTT 1K-A | R@1: 44.2 -> 45.4 | Improved |
| use Token Shift for temporal modeling instead of Prompt Switch (ablation) | CLIP · MSR-VTT 1K-A | R@1: 45.4 -> 43.2 | Hurt |
| use full attention over all frame patches instead of Prompt Switch (ablation) | CLIP · MSR-VTT 1K-A | R@1: 45.4 -> 43.8 | Hurt |
| replace mean-pooling with X-Pool cross-modal fusion on top of Prompt Switch (ablation) | Prompt Switch · MSR-VTT 1K-A | R@1 (text-to-video): 46.1 -> 47.8 | Improved |
| use text-agnostic mean pooling instead of X-Pool fusion (ablation) | Prompt Switch · MSR-VTT 1K-A | ranking latency: 3.9s -> 8.1ms | Improved |
| randomly subsample k=3 of 6 prompt vectors during training (ablation) | Prompt Switch · MSR-VTT 1K-A | retrieval performance | Improved |
| set captioning loss weight lambda to 0.5 instead of 0.3 (ablation) | Prompt Switch · MSR-VTT 1K-A | retrieval performance | Improved |
| use mean-pool Prompt Switch instead of cross-modal methods on LSMDC | Prompt Switch · LSMDC | Meta Sum: 238.9 (X-Pool) vs 228.4 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- X-Pool — X-Pool: Cross-Modal Language-Video Attention for Text-Video Retrieval (CVPR 2022) · better on 3 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DGL: Dynamic Global-Local Prompt Tuning for Text-Video Retrieval (AAAI 2024 · similarity 0.90)
- Holistic Features are almost Sufficient for Text-to-Video Retrieval (CVPR 2024 · similarity 0.89)
- Towards Efficient and Effective Text-to-Video Retrieval with Coarse-to-Fine Visual Representation Learning (AAAI 2024 · similarity 0.89)
- MV-Adapter: Multimodal Video Transfer Learning for Video Text Retrieval (CVPR 2024 · similarity 0.89)
- Mitigating Semantic Collapse in Partially Relevant Video Retrieval (NeurIPS 2025 · similarity 0.89)
- TempMe: Video Temporal Token Merging for Efficient Text-Video Retrieval (ICLR 2025 · similarity 0.89)
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.
