Research map / Video / Video-language understanding
Meta-optimized Angular Margin Contrastive Framework for Video-Language Representation Learning
Thong Thanh Nguyen, Yi Bin, Xiaobao Wu, Xinshuai Dong, Zhiyuan Hu, Khoi M Le, Cong-Duy Nguyen, See Kiong Ng, Anh Tuan Luu
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 LLaVA-generated video-text data (ablation) | MAMA · MSRVTT-QA | accuracy: 45.6 -> 46.4 | Improved |
| use LLaVA instead of BLIP-2 or InstructBLIP for augmentation (ablation) | MAMA · DiDeMo text-to-video retrieval | R@1: 58.1 / 58.5 -> 62.7 | Improved |
| summarize per-frame LLaVA captions with ChatGPT instead of the grid prompt (ablation) | MAMA · DiDeMo text-to-video retrieval | R@1: 62.7 -> 58.7 | Hurt |
| use MLP-parameterized weighting function instead of focal loss, SPL or L2RW (ablation) | MAMA · MSRVTT-QA | accuracy: 45.0 / 45.3 / 45.9 -> 46.4 | Improved |
| use meta-learning instead of joint learning for the weighting MLP (ablation) | MAMA · DiDeMo text-to-video retrieval | R@1: 59.2 -> 62.7 | Improved |
| increase key frames Q up to 12 in a 3x4 grid (ablation) | MAMA · MSRVTT-QA | accuracy: 45.0 -> 46.4 | Improved |
| increase key frames beyond 12 to 16 (ablation) | MAMA · MSRVTT-QA | accuracy: 46.4 -> 45.7 | Hurt |
| use balanced grid (3x4) instead of skewed grid (2x6) (ablation) | MAMA · MSRVTT-QA | accuracy: 45.8 -> 46.4 | Improved |
| add subtractive angular margin contrastive loss with margin around 0.2 | CLIP-ViP · MSRVTT, DiDeMo, ActivityNet retrieval | R@1: 55.7 -> 60.0 on MSRVTT | Improved |
| apply MAMA to VGT backbone for open-ended VideoQA | VGT · MSVD-QA | accuracy: 36.4 -> 37.1 | Improved |
| apply MAMA to VIOLET instead of VGT for Causal-VidQA counterfactual | MAMA · Causal-VidQA (Counterfactual) | accuracy: 68.2 (VGT) vs 59.5 (VIOLET) | Mixed |
Closest papers by meaning
- Distilling Vision-Language Models on Millions of Videos (CVPR 2024 · similarity 0.88)
- Video-Panda: Parameter-efficient Alignment for Encoder-free Video-Language Models (CVPR 2025 · similarity 0.88)
- LanguageBind: Extending Video-Language Pretraining to N-modality by Language-based Semantic Alignment (ICLR 2024 · similarity 0.88)
- BOLT: Boost Large Vision-Language Model Without Training for Long-form Video Understanding (CVPR 2025 · similarity 0.87)
- Needle In A Video Haystack: A Scalable Synthetic Evaluator for Video MLLMs (ICLR 2025 · similarity 0.87)
- TOPA: Extending Large Language Models for Video Understanding via Text-Only Pre-Alignment (NeurIPS 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.
