Research map / Vision-language and multimodal / CLIP and zero-shot
Bidirectional Cross-Modal Knowledge Exploration for Video Recognition With Pre-Trained Vision-Language Models
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 |
|---|---|---|---|
| Replace mean pooling with Video Concept Spotting (ablation) | BIKE video branch · Kinetics-400 | Top-1 accuracy: 76.8 -> 78.5 (+1.7) | Helped |
| Add a temporal Transformer to Video Concept Spotting (ablation) | BIKE video branch · Kinetics-400 | Top-1 accuracy: 78.5 -> 78.7 (+0.2%) | Helped |
| Freeze the category encoder (ablation) | BIKE video branch · Kinetics-400 | Top-1 accuracy: 78.7 -> 78.9 (+0.2%) | Helped |
| Use word embeddings for saliency and [CLS] embedding for classification (ablation) | BIKE Video Concept Spotting · Kinetics-400 | Top-1 accuracy: 74.7 -> 78.5 | Helped |
| Add the manual textual prompt (ablation) | BIKE attributes branch · Kinetics-400 | Top-1 accuracy: 51.2 -> 56.6 | Helped |
| Vary the number of retrieved attributes from 3 to 7 (ablation) | BIKE attributes branch · Kinetics-400 | Top-1 accuracy: 53.4; 56.6; 57.1 | No clear effect |
| Fine-tune the attributes encoder (ablation) | BIKE attributes branch · Kinetics-400 | Top-1 accuracy: 56.6 -> 69.6 | Helped |
| Add the untrained attributes branch to the video branch (ablation) | BIKE video branch · Kinetics-400 | Top-1 accuracy: 78.9 -> 80.0 (+1.1%) | Helped |
| Fine-tune and fuse the attributes branch (ablation) | BIKE video branch · Kinetics-400 | Top-1 accuracy: 78.9 -> 81.4 (+2.5%) | Helped |
| Use ImageNet-1K instead of Kinetics-400 as the attribute lexicon (ablation) | BIKE attributes branch · Kinetics-400 | Top-1 accuracy: 81.4 -> 80.3 | Hurt |
| Add the attributes branch to the video branch (ablation) | BIKE ViT-L/14 multiple-view model · Kinetics-400 | Top-1 accuracy: 87.4 -> 87.4 (+0%) | No clear effect |
| Increase input resolution to 336px (ablation) | BIKE ViT-L · HMDB-51 (split 1) | Top-1 accuracy: 82.2% -> 83.1% | Helped |
Closest papers by meaning
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.87)
- VicTR: Video-conditioned Text Representations for Activity Recognition (CVPR 2024 · similarity 0.87)
- PerceptionLM: Open-Access Data and Models for Detailed Visual Understanding (NeurIPS 2025 · similarity 0.87)
- OST: Refining Text Knowledge with Optimal Spatio-Temporal Descriptor for General Video Recognition (CVPR 2024 · similarity 0.87)
- Leveraging Textual Memory and Key Frame Reasoning for Full Video Understanding Using Off-the-Shelf LLMs and VLMs (Student Abstract) (AAAI 2025 · similarity 0.87)
- Look, Remember and Reason: Grounded Reasoning in Videos with Language Models (ICLR 2024 · 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.