TasksSotADatasetsPapersMethodsSubmitAbout
Papers With Code 2

A community resource for machine learning research: papers, code, benchmarks, and state-of-the-art results.

Explore

Notable BenchmarksAll SotADatasetsPapersMethods

Community

Submit ResultsAbout

Data sourced from the PWC Archive (CC-BY-SA 4.0). Built by the community, for the community.

Papers/Masked Autoencoders Are Scalable Vision Learners

Masked Autoencoders Are Scalable Vision Learners

Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick

2021-11-11CVPR 2022 1Self-Supervised Image ClassificationImage ClassificationSelf-Supervised LearningDomain GeneralizationSemantic SegmentationOut-of-Distribution GeneralizationObject Detection
PaperPDFCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCode(official)CodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCodeCode

Abstract

This paper shows that masked autoencoders (MAE) are scalable self-supervised learners for computer vision. Our MAE approach is simple: we mask random patches of the input image and reconstruct the missing pixels. It is based on two core designs. First, we develop an asymmetric encoder-decoder architecture, with an encoder that operates only on the visible subset of patches (without mask tokens), along with a lightweight decoder that reconstructs the original image from the latent representation and mask tokens. Second, we find that masking a high proportion of the input image, e.g., 75%, yields a nontrivial and meaningful self-supervisory task. Coupling these two designs enables us to train large models efficiently and effectively: we accelerate training (by 3x or more) and improve accuracy. Our scalable approach allows for learning high-capacity models that generalize well: e.g., a vanilla ViT-Huge model achieves the best accuracy (87.8%) among methods that use only ImageNet-1K data. Transfer performance in downstream tasks outperforms supervised pre-training and shows promising scaling behavior.

Results

TaskDatasetMetricValueModel
Domain AdaptationImageNet-RTop-1 Error Rate33.5MAE (ViT-H, 448)
Domain AdaptationImageNet-ATop-1 accuracy %76.7MAE (ViT-H, 448)
Domain AdaptationImageNet-Cmean Corruption Error (mCE)33.8MAE (ViT-H)
Domain AdaptationImageNet-SketchTop-1 accuracy50.9MAE (ViT-H, 448)
Semantic SegmentationImageNet-SmIoU (test)61.2MAE (ViT-B/16, 224x224, SSL+FT, mmseg)
Semantic SegmentationImageNet-SmIoU (val)61.6MAE (ViT-B/16, 224x224, SSL+FT, mmseg)
Semantic SegmentationImageNet-SmIoU (test)60.2MAE (ViT-B/16, 224x224, SSL+FT)
Semantic SegmentationImageNet-SmIoU (val)61MAE (ViT-B/16, 224x224, SSL+FT)
Semantic SegmentationImageNet-SmIoU (test)40.3MAE (ViT-B/16, 224x224, SSL, mmseg)
Semantic SegmentationImageNet-SmIoU (val)40MAE (ViT-B/16, 224x224, SSL, mmseg)
Semantic SegmentationImageNet-SmIoU (test)37MAE (ViT-B/16, 224x224, SSL)
Semantic SegmentationImageNet-SmIoU (val)38.3MAE (ViT-B/16, 224x224, SSL)
Semantic SegmentationADE20KValidation mIoU53.6MAE (ViT-L, UperNet)
Semantic SegmentationADE20KValidation mIoU48.1MAE (ViT-B, UperNet)
Object DetectionCOCO minivalbox AP53.3MAE (ViT-L, Mask R-CNN)
Object DetectionCOCO minivalbox AP50.3MAE (ViT-B, Mask R-CNN)
Image ClassificationiNaturalistTop 1 Accuracy83.4MAE (ViT-H, 448)
Image ClassificationPlaces205Top 1 Accuracy66.8MAE (ViT-H, 448)
Image ClassificationOmniBenchmarkAverage Top-1 Accuracy30.6MAE
Image ClassificationPlaces365-StandardTop 1 Accuracy60.3MAE (ViT-H, 448)
Image ClassificationiNaturalist 2019Top-1 Accuracy88.3MAE (ViT-H, 448)
3DCOCO minivalbox AP53.3MAE (ViT-L, Mask R-CNN)
3DCOCO minivalbox AP50.3MAE (ViT-B, Mask R-CNN)
2D ClassificationCOCO minivalbox AP53.3MAE (ViT-L, Mask R-CNN)
2D ClassificationCOCO minivalbox AP50.3MAE (ViT-B, Mask R-CNN)
2D Object DetectionCOCO minivalbox AP53.3MAE (ViT-L, Mask R-CNN)
2D Object DetectionCOCO minivalbox AP50.3MAE (ViT-B, Mask R-CNN)
Domain GeneralizationImageNet-RTop-1 Error Rate33.5MAE (ViT-H, 448)
Domain GeneralizationImageNet-ATop-1 accuracy %76.7MAE (ViT-H, 448)
Domain GeneralizationImageNet-Cmean Corruption Error (mCE)33.8MAE (ViT-H)
Domain GeneralizationImageNet-SketchTop-1 accuracy50.9MAE (ViT-H, 448)
10-shot image generationImageNet-SmIoU (test)61.2MAE (ViT-B/16, 224x224, SSL+FT, mmseg)
10-shot image generationImageNet-SmIoU (val)61.6MAE (ViT-B/16, 224x224, SSL+FT, mmseg)
10-shot image generationImageNet-SmIoU (test)60.2MAE (ViT-B/16, 224x224, SSL+FT)
10-shot image generationImageNet-SmIoU (val)61MAE (ViT-B/16, 224x224, SSL+FT)
10-shot image generationImageNet-SmIoU (test)40.3MAE (ViT-B/16, 224x224, SSL, mmseg)
10-shot image generationImageNet-SmIoU (val)40MAE (ViT-B/16, 224x224, SSL, mmseg)
10-shot image generationImageNet-SmIoU (test)37MAE (ViT-B/16, 224x224, SSL)
10-shot image generationImageNet-SmIoU (val)38.3MAE (ViT-B/16, 224x224, SSL)
10-shot image generationADE20KValidation mIoU53.6MAE (ViT-L, UperNet)
10-shot image generationADE20KValidation mIoU48.1MAE (ViT-B, UperNet)
16kCOCO minivalbox AP53.3MAE (ViT-L, Mask R-CNN)
16kCOCO minivalbox AP50.3MAE (ViT-B, Mask R-CNN)

Related Papers

SeC: Advancing Complex Video Object Segmentation via Progressive Concept Construction2025-07-21Automatic Classification and Segmentation of Tunnel Cracks Based on Deep Learning and Visual Explanations2025-07-18Adversarial attacks to image classification systems using evolutionary algorithms2025-07-17Efficient Adaptation of Pre-trained Vision Transformer underpinned by Approximately Orthogonal Fine-Tuning Strategy2025-07-17Federated Learning for Commercial Image Sources2025-07-17MUPAX: Multidimensional Problem Agnostic eXplainable AI2025-07-17A Semi-Supervised Learning Method for the Identification of Bad Exposures in Large Imaging Surveys2025-07-17Simulate, Refocus and Ensemble: An Attention-Refocusing Scheme for Domain Generalization2025-07-17