Research map / Video / Action recognition
VATT: Transformers for Multimodal Self-Supervised Learning from Raw Video, Audio and Text
Hassan Akbari, Liangzhe Yuan, Rui Qian, Wei-Hong Chuang, Shih-Fu Chang, Yin Cui, Boqing Gong
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 |
|---|---|---|---|
| Pre-train the vision Transformer with multimodal self-supervision rather than training from scratch (ablation) | VATT vision backbone · ImageNet | top-1 accuracy: 64.7 -> 78.7 | Improved |
| Share one Transformer backbone across video, audio, and text instead of using modality-specific backbones (ablation) | VATT · Kinetics-400 | top-1 accuracy: 79.6 -> 79.9 | No clear effect |
| Share one Transformer backbone across video, audio, and text instead of using modality-specific backbones (ablation) | VATT · AudioSet | mAP: 39.4 -> 39.3 | No clear effect |
| Share one Transformer backbone across video, audio, and text (ablation) | VATT · YouCook2 | R@10: 45.5 -> 40.6 | Hurt |
| Increase DropToken rate from 0% to 75% (ablation) | VATT-MBS · HMDB51 | accuracy: 66.4 -> 62.5 | Hurt |
| Increase DropToken rate from 0% to 75% (ablation) | VATT-MBS · UCF101 | accuracy: 87.6 -> 84.0 | Hurt |
| Use higher-resolution 32×320×320 inputs with 50% DropToken instead of 32×224×224 without dropping tokens (ablation) | VATT video fine-tuning · Kinetics-400 | top-1 accuracy: 79.9 -> 80.2 | Improved |
| Increase sampled frames from 32 to 64 while using 5×16×16 patches (ablation) | VATT video input · UCF101 | accuracy: 87.74 -> 86.57 | Hurt |
| Increase video patch size from 4×16×16 to 8×32×32 (ablation) | VATT video input · UCF101 | accuracy: 87.8 -> 82.68 | Hurt |
| Increase waveform patch size from 128 to 256 (ablation) | VATT audio input · ESC50 | accuracy: 87.75 -> 83.75 | Hurt |
| Replace raw waveform input with spectrogram input (ablation) | VATT audio input · ESC50 | accuracy: 87.75 vs. 88 | No clear effect |
Later papers that compared with this method
4 later accepted papers report VATT as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- Foley-Flow: Coordinated Video-to-Audio Generation with Masked Audio-Visual Alignment and Dynamic Conditional Flows (CVPR 2025 · better on 2 of 2 shared results)
- CAFA: a Controllable Automatic Foley Artist (ICCV 2025 · better on 0 of 1 shared result)
- Language-based Action Concept Spaces Improve Video Self-Supervised Learning (NeurIPS 2023 · better on 1 of 1 shared result)
- Scaling Multimodal Pre-Training via Cross-Modality Gradient Harmonization (NeurIPS 2022 · better on 9 of 12 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Parameter Efficient Multimodal Transformers for Video Representation Learning (ICLR 2021 · similarity 0.89)
- Self-Supervised MultiModal Versatile Networks (NeurIPS 2020 · similarity 0.88)
- ViViT: A Video Vision Transformer (ICCV 2021 · similarity 0.87)
- MJEPA: A Simple and Scalable Joint-Embedding Predictive Architecture for Audio-Visual Learning (ECCV 2026 · similarity 0.87)
- Adapting Short-Term Transformers for Action Detection in Untrimmed Videos (CVPR 2024 · similarity 0.86)
- Unified-IO 2: Scaling Autoregressive Multimodal Models with Vision Language Audio and Action (CVPR 2024 · 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.
