Research map / Video / Video segmentation and tracking
OmniMAE: Single Model Masked Pretraining on Images and Videos
Rohit Girdhar, Alaaeldin El-Nouby, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, Ishan Misra
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 |
|---|---|---|---|
| Jointly pretrain one ViT on images and videos | OmniMAE · Image and video recognition benchmarks | top-1 accuracy | Mixed |
| Increase video masking from 75% to 95% (ablation) | OmniMAE ViT-B · Something-Something-v2 | top-1 accuracy: 67.2 -> 68.6 | Improved |
| Increase image masking from 75% to 90% (ablation) | OmniMAE ViT-B · ImageNet-1K | top-1 accuracy: 82.6 -> 82.8 | No clear effect |
| Use causal instead of random image masking (ablation) | OmniMAE ViT-B · ImageNet-1K | top-1 accuracy: 82.8 -> 82.1 | Hurt |
| Use frame instead of tube video masking (ablation) | OmniMAE ViT-B · Something-Something-v2 | top-1 accuracy: 67.9 -> 65.8 | Hurt |
| Use causal instead of tube video masking (ablation) | OmniMAE ViT-B · Something-Something-v2 | top-1 accuracy: 67.9 -> 64.4 | Hurt |
| Use random instead of tube video masking (ablation) | OmniMAE ViT-B · Something-Something-v2 | top-1 accuracy: 67.9 -> 67.8 | No clear effect |
| Share one decoder across image and video (ablation) | OmniMAE ViT-B · ImageNet-1K | top-1 accuracy: 82.6 -> 82.8 | Improved |
| Use an 8-layer rather than 4-layer decoder (ablation) | OmniMAE ViT-B · ImageNet-1K | top-1 accuracy: 82.8 -> 82.8 | No clear effect |
| Oversample video data relative to image data (ablation) | OmniMAE ViT-B · Something-Something-v2 | top-1 accuracy | Improved |
| Replicate samples within a mini-batch using different crops and masks (ablation) | OmniMAE · Video training | training speed: 20% faster training | Improved |
| Use video clips rather than individual video frames in joint pretraining (ablation) | OmniMAE ViT-B · Something-Something-v2 | top-1 accuracy: 64.2 -> 69.0 | Improved |
Closest papers by meaning
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · similarity 0.90)
- Asymmetric Masked Distillation for Pre-Training Small Foundation Models (CVPR 2024 · similarity 0.89)
- VideoMAC: Video Masked Autoencoders Meet ConvNets (CVPR 2024 · similarity 0.89)
- EVEREST: Efficient Masked Video Autoencoder by Removing Redundant Spatiotemporal Tokens (ICML 2024 · similarity 0.88)
- 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)
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.
