Research map / Vision-language and multimodal / Cross-modal learning
Audio-Visual Instance Discrimination with Cross-Modal Agreement
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 cross-modal instance discrimination instead of within-modal instance discrimination (ablation) | AVID · Kinetics and ESC linear probing | top-1 accuracy: 32.08 -> 39.95 on Kinetics; 72.64 -> 75.05 on ESC | Helped |
| Add within-modal instance discrimination to Cross-AVID (ablation) | AVID · Kinetics and ESC linear probing | top-1 accuracy: 39.95 -> 33.04 on Kinetics; 75.05 -> 71.77 on ESC | Hurt |
| Expand positives using audio similarity only (ablation) | Cross-AVID · Kinetics linear probing | top-1 accuracy (Best): 39.95 -> 39.62 | Hurt |
| Expand positives using video similarity only (ablation) | Cross-AVID · Kinetics linear probing | top-1 accuracy (Best): 39.95 -> 40.17 | Helped |
| Expand positives using similarity in either audio or video (ablation) | Cross-AVID · Kinetics linear probing | top-1 accuracy (Best): 39.95 -> 40.58 | Helped |
| Choose positives by agreement in both audio and video spaces (ablation) | Cross-AVID · Kinetics linear probing | top-1 accuracy (Best): 39.95 -> 41.11 | Helped |
| Add within-modal positive discrimination using CMA-selected positives | Cross-AVID · Kinetics linear probing, Kinetics pretraining | Clip@1: 33.3 -> 35.1 | Helped |
| Add within-modal positive discrimination using CMA-selected positives | Cross-AVID · UCF-101 and HMDB-51 fine-tuning, Kinetics pretraining, 32 frames | top-1 accuracy: 86.9 -> 87.5 on UCF; 59.9 -> 60.8 on HMDB | Helped |
| Add within-modal positive discrimination using CMA-selected positives | Cross-AVID · ESC-50 linear probing, Audioset pretraining | top-1 accuracy: 89.2 -> 89.1 | Hurt |
| Train Cross-AVID for 400 rather than 200 epochs (ablation) | Cross-AVID · Kinetics linear probing | top-1 accuracy (Best): 39.87 -> 39.95 | No clear effect |
Closest papers by meaning
- Robust Audio-Visual Instance Discrimination (CVPR 2021 · similarity 0.96)
- Connect, Collapse, Corrupt: Learning Cross-Modal Tasks with Uni-Modal Data (ICLR 2024 · similarity 0.88)
- Looking Similar Sounding Different: Leveraging Counterfactual Cross-Modal Pairs for Audiovisual Representation Learning (CVPR 2024 · similarity 0.88)
- XKD: Cross-Modal Knowledge Distillation with Domain Alignment for Video Representation Learning (AAAI 2024 · similarity 0.88)
- CrossMAE: Cross-Modality Masked Autoencoders for Region-Aware Audio-Visual Pre-Training (CVPR 2024 · similarity 0.88)
- EquiAV: Leveraging Equivariance for Audio-Visual Contrastive Learning (ICML 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.