Research map / Vision-language and multimodal / Cross-modal learning
Mettle: Meta-Token Learning for Memory-Efficient Audio-Visual Adaptation
Jinxing Zhou, Zhihui Li, Yongqiang Yu, Yanghao Zhou, Ruohao Guo, Guangyao Li, Yuxin Mao, Mingfei Han, Xiaojun Chang, Meng Wang
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 task-specific meta-token branch (TS) (ablation) | LCD · AVE (Swin-V2-L shared) | accuracy: 77.3 -> 82.3 | Improved |
| add linear pretrained-knowledge branch (DP) (ablation) | LCD · AVE (Swin-V2-L shared) | accuracy: 80.5 -> 82.3 | Improved |
| add cross-modal injection to MTI (ablation) | MTI · AVSBench MS3 (Swin-V2-L shared) | mIoU: 49.9 -> 53.7 | Improved |
| add intra-modal injection to MTI (ablation) | MTI · AVSBench MS3 (Swin-V2-L shared) | mIoU: 52.3 -> 53.7 | Improved |
| increase meta-tokens from 1 to 4 per modality (ablation) | Mettle · AVE (Swin-V2-L shared) | accuracy: 82.3 -> 80.8 | Hurt |
| use hierarchical meta-token numbers {8,4,2,1} (ablation) | Mettle · AVE (Swin-V2-L shared) | accuracy: 82.3 -> 80.7 | Hurt |
| increase distillation steps to Ra=4, Rv=4 (ablation) | Mettle · AVE (Swin-V2-L shared) | accuracy: 82.3 -> 81.6 | Hurt |
| replace linear branch with average pooling (ablation) | LCD · AVE (Swin-V2-L + HTS-AT) | accuracy: 83.3 -> 83.2 | No clear effect |
| replace linear branch with average pooling (ablation) | LCD · AVSBench MS3 (Swin-V2-L + HTS-AT) | mIoU: 55.1 -> 53.8 | Hurt |
| add linear branch instead of removing it (ablation) | LCD · LLP AVVP (Swin-V2-L + HTS-AT) | average F1: 53.6 -> 54.7 | Improved |
| reverse MTI injection order to intra-first (ablation) | MTI · AVSBench MS3 (Swin-V2-L shared) | mIoU: 53.7 -> 53.3 | No clear effect |
| apply LCD at top layer instead of bottom layer for MTI (ablation) | MTI · AVSBench MS3 (Swin-V2-L + HTS-AT, MTI placement) | mIoU: 53.6 -> 55.1 | Improved |
Closest papers by meaning
- CAV-MAE Sync: Improving Contrastive Audio-Visual Mask Autoencoders via Fine-Grained Alignment (CVPR 2025 · similarity 0.88)
- From Vision to Audio and Beyond: A Unified Model for Audio-Visual Representation and Generation (ICML 2024 · similarity 0.88)
- Mixtures of Experts for Audio-Visual Learning (NeurIPS 2024 · similarity 0.87)
- Prompting Segmentation with Sound Is Generalizable Audio-Visual Source Localizer (AAAI 2024 · similarity 0.87)
- Separating the "Chirp" from the "Chat": Self-supervised Visual Grounding of Sound and Language (CVPR 2024 · similarity 0.87)
- EquiAV: Leveraging Equivariance for Audio-Visual Contrastive Learning (ICML 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.
