AI Paper Atlas

Research map / 3D vision / Human pose estimation

TokenPose: Learning Keypoint Tokens for Human Pose Estimation

Yanjie Li, Shoukui Zhang, Zhicheng Wang, Sen Yang, Wankou Yang, Shu-Tao Xia, Erjin Zhou

ICCV 2021 · 3D vision › Human pose estimation · approach: Heatmap and coordinate pose encoding · 472 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 2D sine position embedding (vs none) (ablation)TokenPose-S-v1 · COCO val2017AP: 67.0 -> 72.5Improved
use 2D sine instead of learnable position embedding (ablation)TokenPose-S-v1 · COCO val2017AP: 71.4 -> 72.5Improved
add learnable position embedding (vs none) (ablation)TokenPose-S-v1 · COCO val2017AP: 67.0 -> 71.4Improved
add keypoint token fusion (ablation)TokenPose-L+/D12 · COCO val2017AP: 75.3 -> 75.5Improved
add keypoint token fusion (ablation)TokenPose-S · COCO val2017AP: 73.5 -> 72.6Hurt
increase depth from 12 to 24 layers (ablation)TokenPose-L · COCO val2017AP: 75.3 -> 75.8Improved
increase embedding width from 192 to 384 (12 layers) (ablation)TokenPose-L · COCO val2017AP: 75.3 -> 75.5Improved
replace SimpleBaseline-Res50 CNN with TokenPose-S-v1SimpleBaseline-Res50 · COCO val2017AP: 70.4 -> 72.5Improved
replace stem-net with HRNet-W32 backbone feeding tokens (TokenPose-B vs HRNet-W32)HRNet-W32 · COCO val2017AP: 74.4 -> 74.7No clear effect
use pure Transformer without any convolutionTokenPose-T · COCO val2017AP: 65.6 vs 72.5Hurt
increase input resolution from 256x192 to 384x288TokenPose-L/D24 · COCO test-dev2017AP: 75.1 -> 75.9Improved
use patch size 2x2 instead of 4x3 (S-v2 vs S-v1)TokenPose-S · COCO val2017AP: 72.5 -> 73.5Improved

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