CrowdPose.md

CrowdPose: Efficient Crowded Scenes Pose Estimation and A New Benchmark (accepted to CVPR2019)

Introduction

Our proposed method surpasses the state-of-the-art methods on CrowdPose dataset by 5 mAP and results on MSCOCO dataset demonstrate the generalization ability of our method (comparatively 0.8 mAP higher). Images in our proposed CrowdPose dataset have a uniform distribution of Crowd Index among [0, 1].

Code

We provide evaluation tools for CrowdPose dataset. Our evaluation tools is developed based on @cocodataset/cocoapi. The source code of our model is integrated into AlphaPose.

Quick Start

Run with matching option to use the matching algorithm in CrowdPose.

• Input dir: Run AlphaPose for all images in a folder with:

python3 demo.py --indir ${img_directory} --outdir examples/res --matching

Dataset

Train + Validation + Test Images (Google Drive)

Annotations (Google Drive)

Results

Results on CrowdPose Validation:

Compare with state-of-the-art methods

Method	AP @0.5:0.95	AP @0.5	AP @0.75	AR @0.5:0.95	AR @0.5	AR @0.75
Detectron (Mask R-CNN)	57.2	83.5	60.3	65.9	89.3	69.4
Simple Pose (Xiao et al.)	60.8	81.4	65.7	67.3	86.3	71.8
Ours	66.0	84.2	71.5	72.7	89.5	77.5

Compare with open-source systems

Method	AP @Easy	AP @Medium	AP @Hard	FPS
OpenPose (CMU-Pose)	62.7	48.7	32.3	5.3
Detectron (Mask R-CNN)	69.4	57.9	45.8	2.9
Ours	75.5	66.3	57.4	10.1

Results on MSCOCO Validation:

Method	AP @0.5:0.95	AR @0.5:0.95
Detectron (Mask R-CNN)	64.8	71.1
Simple Pose (Xiao et al.)	69.8	74.1
AlphaPose	70.9	76.4

Contributors

CrowdPose is authored by Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu.