Research map / Video / Action recognition
Vita-CLIP: Video and Text Adaptive CLIP via Multimodal Prompting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 global video-level prompts (Mv=8) on the vision side (ablation) | Vita-CLIP B/16 with CSC text prompts · Kinetics-400 | Top-1: 73.00 -> 77.83 | Helped |
| add local frame-level prompts conditioned on CLS tokens (ablation) | Vita-CLIP B/16 with CSC and G · Kinetics-400 | Top-1: 77.83 -> 79.16 | Helped |
| add summary token (ablation) | Vita-CLIP B/16 with CSC, G, and L · Kinetics-400 | Top-1: 79.16 -> 80.51 | Helped |
| use class-specific context (CSC) instead of unified context (UC) on text side (ablation) | Vita-CLIP text encoder · Kinetics-400 | Top-1: CSC 80.51 vs UC 79.33 at 32 prompts | Helped |
| increase text context prompts beyond 8 (Mc=16 or 32) with CSC (ablation) | Vita-CLIP text encoder · Kinetics-400 | Top-1: 80.51 (Mc=8) vs 80.46 (16) vs 80.56 (32) | No clear effect |
| increase number of global video-level prompts beyond 8 (ablation) | Vita-CLIP video prompts · Kinetics-400 | Top-1: 80.51 (Mv=8) to 80.92 (Mv=25) | No clear effect |
| frozen CLIP backbone with prompts instead of full finetuning (vs X-CLIP) | Vita-CLIP B/16 · HMDB51 and UCF101 zero-shot | Top-1 accuracy: 48.6 vs 44.6 (HMDB51); 75.0 vs 72.0 (UCF101) | Helped |
| frozen backbone with prompts (vs finetuned X-CLIP) under the same single training | Vita-CLIP B/16 on Kinetics-400 supervised · Kinetics-400 | Top-1: 80.5 (8 frames, 1x1) vs X-CLIP 82.3 (8 frames, 1x1) | Mixed |
| frozen backbone with prompts on SSv2 supervised | Vita-CLIP B/16 · Something-Something-V2 | Top-1: 48.7 vs fine-tuned methods 48.8 to 69.6 | Mixed |
Closest papers by meaning
- ViLT-CLIP: Video and Language Tuning CLIP with Multimodal Prompt Learning and Scenario-Guided Optimization (AAAI 2024 · similarity 0.93)
- A Hard-to-Beat Baseline for Training-free CLIP-based Adaptation (ICLR 2024 · similarity 0.89)
- Enhancing Zero-Shot Vision Models by Label-Free Prompt Distribution Learning and Bias Correcting (NeurIPS 2024 · similarity 0.89)
- Context-Aware Multimodal Pretraining (CVPR 2025 · similarity 0.88)
- Understanding Transferable Representation Learning and Zero-shot Transfer in CLIP (ICLR 2024 · similarity 0.88)
- PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 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.