Research map / Video / Video segmentation and tracking
Masked Feature Prediction for Self-Supervised Visual Pre-Training
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 |
|---|---|---|---|
| Replace RGB pixel targets with HOG feature targets | MaskFeat · Kinetics-400 and ImageNet-1K feature-target studies | top-1 accuracy: 80.7 -> 82.2; 82.5 -> 83.6 | Helped |
| Use supervised features or class pseudo-labels as prediction targets | MaskFeat · ImageNet-1K and Kinetics-400 | top-1 accuracy: 81.9; 81.9 | Hurt |
| Add cube masking that combines spatial and temporal blocks (ablation) | MaskFeat video pre-training · Kinetics-400 | top-1 accuracy: 81.0 -> 82.2; 81.9 -> 82.2 | Helped |
| Predict HOG for the entire masked cube rather than its center patch (ablation) | MaskFeat video pre-training · Kinetics-400 | top-1 accuracy: 82.2 -> 82.0 | No clear effect |
| Increase the masking ratio from 40% to 60% or 80% (ablation) | MaskFeat video pre-training · Kinetics-400 | top-1 accuracy: 82.2 -> 82.2; 82.2 -> 82.0 | No clear effect |
| Reduce the video masking ratio from 40% to 20% (ablation) | MaskFeat video pre-training · Kinetics-400 | top-1 accuracy: 82.2 -> 81.9 | Hurt |
| Extend pre-training from 300 to 800 epochs (ablation) | MaskFeat MViT-S · Kinetics-400 | top-1 accuracy: 82.2 -> 82.0 | Hurt |
| Extend pre-training from 300 to 800 epochs (ablation) | MaskFeat MViT-L · Kinetics-400 | top-1 accuracy: 83.1 -> 84.3 | Helped |
| Add local contrast normalization to HOG targets (ablation) | MaskFeat image pre-training · ImageNet-1K | top-1 accuracy: 82.2 -> 83.6 | Helped |
| Use RGB color channels rather than grayscale HOG (ablation) | MaskFeat image pre-training · ImageNet-1K | top-1 accuracy: 83.2 -> 83.6 | Helped |
| Combine pixel and HOG prediction targets with separate heads (ablation) | MaskFeat image pre-training · ImageNet-1K | top-1 accuracy: 83.6 -> 82.3 | Hurt |
| Increase image masking from 40% to 60% or 80% (ablation) | MaskFeat image pre-training · ImageNet-1K | top-1 accuracy: 83.6 -> 83.1; 83.6 -> 82.5 | Hurt |
Closest papers by meaning
- Asymmetric Masked Distillation for Pre-Training Small Foundation Models (CVPR 2024 · similarity 0.87)
- From Prototypes to General Distributions: An Efficient Curriculum for Masked Image Modeling (CVPR 2025 · similarity 0.86)
- VideoMAC: Video Masked Autoencoders Meet ConvNets (CVPR 2024 · similarity 0.85)
- Self-Organizing Visual Prototypes for Non-Parametric Representation Learning (ICML 2025 · similarity 0.85)
- Video Language Model Pretraining with Spatio-temporal Masking (CVPR 2025 · similarity 0.85)
- EVEREST: Efficient Masked Video Autoencoder by Removing Redundant Spatiotemporal Tokens (ICML 2024 · similarity 0.85)
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.