AI Paper Atlas

Research map / Detection and segmentation / CNN features

Unleashing Vanilla Vision Transformer with Masked Image Modeling for Object Detection

Yuxin Fang, Shusheng Yang, Shijie Wang, Yixiao Ge, Ying Shan, Xinggang Wang

ICCV 2023 · Detection and segmentation › CNN features · approach: Plain and high-resolution ViTs · 66 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
randomly sample 50% of input embeddings for ViT encoder during training instead of 100% (ablation)MimDet-Base · COCO val2017APbox: 51.8 -> 51.5No clear effect
reduce training sampling ratio from 50% to 25% (ablation)MimDet-Base · COCO val2017APbox: 51.5 -> 49.9Hurt
use random sampling instead of grid sampling (ablation)MimDet-Base · COCO val2017APbox: 48.7 -> 51.5Improved
add a short full-input fine-tuning stage after sampled fine-tuning (ablation)MimDet-Base · COCO val2017APbox: 51.5 -> 51.4No clear effect
infer with full input instead of ensembling sampled inferences (ablation)MimDet-Base · COCO val2017APbox: 51.0 (8x50%) -> 51.5 (full)Improved
replace naive ConvStem with ConvNeXt early stage (pretrained or not) (ablation)MimDet-Base · COCO val2017APbox: 51.5 -> 51.3 / 51.4No clear effect
fill unsampled positions with ConvStem features instead of [MASK] tokens (ablation)MimDet-Base · COCO val2017APbox: 51.5 -> 51.7Improved
increase decoder layers from 1 to 4 (ablation)MimDet-Base · COCO val2017APbox: 49.9 -> 51.5Improved
increase decoder layers from 4 to 8 (ablation)MimDet-Base · COCO val2017APbox: 51.5 -> 51.6No clear effect
take FPN high-resolution inputs from ConvStem instead of ViT intermediate features (ablation)MimDet-Base · COCO val2017APbox: 48.9 -> 49.5Improved
use random sampled partial input instead of window partitioned input (ablation)MimDet-Base (1 decoder layer) · COCO val2017APbox: 49.5 -> 49.9Improved
use MimDet instead of prior adapted vanilla ViT detectorMAE ViT-Base · COCO val2017APbox: 50.3 -> 51.7Improved

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