AI Paper Atlas

Research map / Detection and segmentation / CNN features

RevColV2: Exploring Disentangled Representations in Masked Image Modeling

Qi Han, Yuxuan Cai, Xiangyu Zhang

NeurIPS 2023 · Detection and segmentation › CNN features · approach: Convolution-transformer backbones · 6 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
keep top-down decoder columns during fine-tuning (vs encoder only) (ablation)RevColV2-B · ImageNet-1Ktop-1 accuracy: 83.8 -> 84.7Improved
keep decoder, matching params/FLOPs for encoder-only variant (ablation)RevColV2-B · ImageNet-1Ktop-1 accuracy: 83.9 -> 84.7Improved
keep vanilla MAE decoder in fine-tuning (ablation)ViT + MAE · ImageNet-1Ktop-1 accuracy: 83.6 -> 83.2Hurt
randomly initialize decoder, use only pre-trained encoder (ablation)RevColV2-B · ImageNet-1Ktop-1 accuracy: 84.7 -> 84.4Hurt
randomly initialize decoder, use only pre-trained encoder (ablation)RevColV2-B · ADE20KmIoU: -0.6Hurt
increase decoder column depth from 4 to 8 (ablation)RevColV2-B · ImageNet-1Ktop-1 accuracy: 84.1 -> 83.6Hurt
increase decoder column depth from 4 to 12 (ablation)RevColV2-B · ImageNet-1Ktop-1 accuracy: 84.1 -> 82.9Hurt
increase decoder column depth from 4 to 8 (ablation)RevColV2-L · ImageNet-1Ktop-1 accuracy: 85.2 -> 84.9Hurt
pre-train 1600 epochs instead of 300 (ablation)RevColV2-B · ImageNet-1Ktop-1 accuracy: 84.1 -> 84.7Improved
pre-train 1600 epochs instead of 300 (ablation)RevColV2-L · ImageNet-1Ktop-1 accuracy: 85.2 -> 86.3Improved
replace reversible column decoder with MAE ViT-block decoder (ablation)RevColV2-B · ImageNet-1Ktop-1 accuracy: 84.1 -> 83.5Hurt
replace reversible column decoder with UNet-like single column (ablation)RevColV2-B · ImageNet-1Ktop-1 accuracy: 84.1 -> 83.4Hurt

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