AI Paper Atlas

Research map / Label-efficient and robust learning / Self-supervised and contrastive

Context Autoencoder for Self-supervised Representation Learning

Xiaokang Chen, Mingyu Ding, Xiaodi Wang, Ying Xin, Shentong Mo, Yunhao Wang, Shumin Han, Ping Luo, Gang Zeng, Jingdong Wang

IJCV 2023 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Masked autoencoder pretraining · 458 citations (OpenAlex)

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 changedWhereMeasuredEffect
add decoder and alignment constraint (ablation)CAE ViT-B 300 ep · ImageNet-1K attentive probingtop-1 accuracy: 71.2 -> 73.8Improved
add alignment on top of decoder (ablation)CAE ViT-B 300 ep · ImageNet-1K attentive probingtop-1 accuracy: 72.7 -> 73.8Improved
add decoder on top of alignment (ablation)CAE ViT-B 300 ep · ImageNet-1K attentive probingtop-1 accuracy: 71.5 -> 73.8Improved
add decoder and alignment (ablation)CAE ViT-B 300 ep · ADE20KmIoU: 47.0 -> 48.3Improved
use 50% instead of 40% mask ratio (ablation)CAE ViT-B 300 ep · ADE20KmIoU: 47.2 -> 48.3Improved
raise mask ratio from 50% to 60% (ablation)CAE ViT-B 300 ep · ImageNet-1K linear probing / ADE20Kaccuracy / mIoU: 64.1 -> 64.8 (LIN), 48.3 -> 48.1 (ADE)Mixed
use 4 layers instead of 1 in regressor and decoder (ablation)CAE · ImageNet-1K attentive probingtop-1 accuracy: 67.5 -> 73.8Improved
use 5 layers instead of 4 in regressor and decoder (ablation)CAE · ImageNet-1K attentive probingtop-1 accuracy: 73.8 -> 73.7No clear effect
set loss tradeoff lambda to 2 instead of 1 (ablation)CAE · ImageNet-1K linear probingtop-1 accuracy: 63.4 -> 64.1Improved
use RGB pixel targets instead of DALL-E tokenizer (ablation)CAE ViT-B 1600 ep · ADE20KmIoU: 50.2 -> 50.4No clear effect
use d-VAE tokenizer (ImageNet-1K) instead of DALL-E tokenizer (ablation)CAE ViT-B 1600 ep · ADE20KmIoU: 50.2 -> 50.1No clear effect
pretrain 1600 instead of 300 epochsCAE ViT-B · ADE20KmIoU: 48.3 -> 50.2Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy