Research map / Video / Video-language understanding
Less Is More: ClipBERT for Video-and-Language Learning via Sparse Sampling
Check 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 |
|---|---|---|---|
| Use sparse random sampling of 4 clips rather than dense uniform sampling of 16 clips (ablation) | ClipBERT · MSRVTT retrieval local val | R1: 15.5 -> 15.7 | Helped |
| Use sparse random sampling of 4 clips rather than dense uniform sampling of 16 clips (ablation) | ClipBERT · MSRVTT-QA val | accuracy: 35.88 -> 36.67 | Helped |
| Add a second frame per clip and mean-pool frame features (ablation) | ClipBERT · MSRVTT retrieval local val | R1: 10.2 -> 11.3 | Helped |
| Replace mean pooling with Conv3D for temporal fusion (ablation) | ClipBERT · MSRVTT retrieval local val | R1: 11.3 -> 8.7 | Hurt |
| Replace mean pooling with Conv(2+1)D for temporal fusion (ablation) | ClipBERT · MSRVTT retrieval local val | R1: 11.3 -> 7.3 | Hurt |
| Increase training clips from 1 to 2 using LogSumExp aggregation (ablation) | ClipBERT · MSRVTT retrieval local val | R1: 12.7 -> 15.5 | Helped |
| Use LogSumExp rather than mean pooling for clip-score aggregation (ablation) | ClipBERT · MSRVTT retrieval local val | R1: 13.3 -> 15.5 | Helped |
| Use LogSumExp rather than max pooling for clip-score aggregation (ablation) | ClipBERT · MSRVTT retrieval local val | R1: 8.5 -> 15.5 | Helped |
| Fine-tune the vision encoder instead of freezing it (ablation) | ClipBERT · MSRVTT retrieval local val | R1: 9.0 -> 10.2 | Helped |
| Fine-tune both vision and language encoders instead of freezing both (ablation) | ClipBERT · MSRVTT retrieval local val | R1: 8.0 -> 10.2 | Helped |
| Use image-text pre-training rather than random initialization (ablation) | ClipBERT · MSRVTT retrieval local val | R1: 0.3 -> 10.2 | Helped |
| Use image-text pre-training rather than ImageNet CNN initialization (ablation) | ClipBERT · MSRVTT retrieval local val | R1: 7.2 -> 10.2 | Helped |
Closest papers by meaning
- Frame-Voyager: Learning to Query Frames for Video Large Language Models (ICLR 2025 · similarity 0.88)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.87)
- Vript: A Video Is Worth Thousands of Words (NeurIPS 2024 · similarity 0.87)
- VideoPrism: A Foundational Visual Encoder for Video Understanding (ICML 2024 · similarity 0.87)
- Distilling Vision-Language Models on Millions of Videos (CVPR 2024 · similarity 0.87)
- Language-aware Visual Semantic Distillation for Video Question Answering (CVPR 2024 · similarity 0.87)
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.