Research map / Video / Action recognition
MeMViT: Memory-Augmented Multiscale Vision Transformer for Efficient Long-Term Video Recognition
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 |
|---|---|---|---|
| Add cached temporal memory (ablation) | MViT-16, 16×4 · AVA | mAP: 27.0 -> 29.3 | Helped |
| Use memory length 1 instead of no memory (ablation) | MViT · AVA | mAP: 27.0 -> 28.7 | Helped |
| Increase memory length from 2 to 3 (ablation) | MeMViT · AVA | mAP: 29.3 -> 29.2 | No clear effect |
| Increase memory length from 2 to 4 (ablation) | MeMViT · AVA | mAP: 29.3 -> 28.8 | Hurt |
| Add learned memory compression with factor 4×2×2 (ablation) | MeMViT · AVA | mAP: 28.9 -> 29.3 | Helped |
| Increase compression to 4×4×4 rather than 4×2×2 (ablation) | MeMViT · AVA | mAP: 29.3 -> 28.6 | Hurt |
| Augment 50% of attention layers uniformly rather than all layers (ablation) | MeMViT · AVA | mAP: 29.1 -> 29.3 | No clear effect |
| Concentrate memory augmentation in early rather than uniformly distributed layers (ablation) | MeMViT · AVA | mAP: 29.3 -> 28.6 | Hurt |
| Concentrate memory augmentation in middle rather than uniformly distributed layers (ablation) | MeMViT · AVA | mAP: 29.3 -> 28.7 | Hurt |
| Concentrate memory augmentation in late rather than uniformly distributed layers (ablation) | MeMViT · AVA | mAP: 29.3 -> 29.1 | No clear effect |
| Replace absolute positional embeddings with relative positional embeddings (ablation) | MViT-B, 16×4 · AVA | mAP: 24.5 -> 25.4 | Helped |
| Pool features before applying linear projections (ablation) | MViT-B, 16×4 with relative positional embeddings · AVA | mAP: 25.4 -> 25.5 | No clear effect |
Closest papers by meaning
- Memory Consolidation Enables Long-Context Video Understanding (ICML 2024 · similarity 0.89)
- Towards Long-Form Video Understanding (CVPR 2021 · similarity 0.89)
- Exploring Temporally Dynamic Data Augmentation for Video Recognition (ICLR 2023 · similarity 0.87)
- Adapting Short-Term Transformers for Action Detection in Untrimmed Videos (CVPR 2024 · similarity 0.87)
- MA-LMM: Memory-Augmented Large Multimodal Model for Long-Term Video Understanding (CVPR 2024 · similarity 0.85)
- Enhancing Temporal Understanding in Video-LLMs through Stacked Temporal Attention in Vision Encoders (NeurIPS 2025 · similarity 0.85)
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.