Research map / Video / Video segmentation and tracking
AdaMAE: Adaptive Masking for Efficient Spatiotemporal Learning With Masked Autoencoders
Read the paperCheck 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 |
|---|---|---|---|
| Replace random patch or tube masking with adaptive token sampling at 95% masking (ablation) | VideoMAE-style masked autoencoder with ViT-B · SSv2 | top-1 accuracy: 68.3 -> 70.04; 69.3 -> 70.04 | Helped |
| Use adaptive token sampling at 95% masking instead of 90% masking (ablation) | AdaMAE · SSv2 | top-1 accuracy: 69.55 -> 70.04 | Helped |
| Use 98% masking instead of the 95% default (ablation) | AdaMAE · SSv2 | top-1 accuracy: 70.04 -> 68.85 | Hurt |
| Use 85% or 80% masking instead of the 95% default (ablation) | AdaMAE · SSv2 | top-1 accuracy: 70.04 -> 68.06; 70.04 -> 66.96 | Hurt |
| Increase decoder depth from 1 to 4 blocks (ablation) | AdaMAE · SSv2 | top-1 accuracy: 68.86 -> 70.04 | Helped |
| Increase decoder depth from 4 to 8 blocks (ablation) | AdaMAE · SSv2 | top-1 accuracy: 70.04 -> 69.97 | Hurt |
| Replace normalized MSE with normalized L1 loss (ablation) | AdaMAE · SSv2 | top-1 accuracy: 70.04 -> 69.12 | Hurt |
| Remove local patch normalization from MSE reconstruction loss (ablation) | AdaMAE · SSv2 | top-1 accuracy: 70.04 -> 68.87 | Hurt |
| Increase pre-training from 200 to 800 epochs (ablation) | AdaMAE · SSv2 | top-1 accuracy: 66.42 -> 70.04 | Helped |
| Replace the MLP sampling network with a single MHA sampler (ablation) | AdaMAE · SSv2 | top-1 accuracy: 65.41 -> 70.04 | Helped |
| Reduce the single MHA sampler embedding dimension from 768 to 384 (ablation) | AdaMAE · SSv2 | top-1 accuracy: 70.04 -> 69.56 | Hurt |
| Increase sampler depth to 2 MHA blocks at dimension 384 (ablation) | AdaMAE · SSv2 | top-1 accuracy: 70.04 -> 70.03 | No clear effect |
Closest papers by meaning
- Extending Video Masked Autoencoders to 128 frames (NeurIPS 2024 · similarity 0.90)
- EVEREST: Efficient Masked Video Autoencoder by Removing Redundant Spatiotemporal Tokens (ICML 2024 · 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 0.87)
- Video Language Model Pretraining with Spatio-temporal Masking (CVPR 2025 · similarity 0.87)
- The Dynamic Duo of Collaborative Masking and Target for Advanced Masked Autoencoder Learning (AAAI 2025 · 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.