Research map / Label-efficient and robust learning / Self-supervised and contrastive
Audiovisual Masked Autoencoders
Mariana-Iuliana Georgescu, Eduardo Fonseca, Radu Tudor Ionescu, Mario Lucic, Cordelia Schmid, Anurag Arnab
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 |
|---|---|---|---|
| pretrain jointly on audio and video instead of audio-only MAE, for audiovisual finetuning (ablation) | Audiovisual MAE · VGGSound audiovisual (ViT-Base) | top-1 accuracy: 58.3 -> 63.5 | Improved |
| pretrain jointly instead of video-only MAE, for audiovisual finetuning (ablation) | Audiovisual MAE · VGGSound audiovisual (ViT-Base) | top-1 accuracy: 62.1 -> 63.5 | Improved |
| audiovisual pretraining for unimodal finetuning (ablation) | Audiovisual MAE · VGGSound audio-only / video-only | top-1 accuracy: 55.8 vs 55.7 audio; 48.5 vs 49.3 video | No clear effect |
| use joint reconstruction objective instead of modality inpainting (ablation) | Audiovisual MAE · VGGSound audiovisual | top-1 accuracy: 58.4 -> 62.2 | Improved |
| use mid-fusion or separate encoder instead of early fusion or shared (ablation) | Audiovisual MAE · VGGSound audiovisual | top-1 accuracy: 62.2 -> 63.5 | Improved |
| use shared decoder instead of early or separate decoder (ablation) | Audiovisual MAE · VGGSound audiovisual | top-1 accuracy: 63.3 -> 63.5 | Improved |
| scale backbone from ViT-Base to ViT-Large with self-supervised pretraining (ablation) | Audiovisual MAE · VGGSound audiovisual | top-1 accuracy: 64.2 -> 65.0 | Improved |
| scale backbone to ViT-Large with supervised ImageNet-21K pretraining (ablation) | MBT · VGGSound audiovisual | top-1 accuracy: 64.1 -> 61.4 | Hurt |
| pretrain longer on VGGSound (800 to 1200 epochs) (ablation) | Audiovisual MAE · VGGSound audiovisual | top-1 accuracy: 65.0 -> 64.9 | No clear effect |
| pretrain longer (200 to 1200 epochs) (ablation) | Audiovisual MAE · Epic Kitchens audiovisual | action accuracy: 41.8 -> 46.0 | Improved |
| pretrain on VGGSound instead of the larger AudioSet (ablation) | Audiovisual MAE · Epic Kitchens audiovisual | action accuracy: 43.5 -> 45.5 | Improved |
| joint audiovisual pretraining instead of two separate unimodal MAEs initialising MBT (ablation) | Audiovisual MAE · VGGSound audiovisual | top-1 accuracy: 63.3 -> 64.2 | Improved |
Closest papers by meaning
- Contrastive Audio-Visual Masked Autoencoder (ICLR 2023 · similarity 0.89)
- Understanding and Enhancing Mask-Based Pretraining towards Universal Representations (NeurIPS 2025 · similarity 0.88)
- Exploring Target Representations for Masked Autoencoders (ICLR 2024 · similarity 0.88)
- Extending Video Masked Autoencoders to 128 frames (NeurIPS 2024 · similarity 0.88)
- Video Language Model Pretraining with Spatio-temporal Masking (CVPR 2025 · similarity 0.88)
- CAV-MAE Sync: Improving Contrastive Audio-Visual Mask Autoencoders via Fine-Grained Alignment (CVPR 2025 · 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.
