Pose-Group / DCPose

This is an official implementation of our CVPR 2021 paper "Deep Dual Consecutive Network for Human Pose Estimation" (https://openaccess.thecvf.com/content/CVPR2021/papers/Liu_Deep_Dual_Consecutive_Network_for_Human_Pose_Estimation_CVPR_2021_paper.pdf)

Deep Dual Consecutive Network for Human Pose Estimation （CVPR2021）

Introduction

This is the official code of Deep Dual Consecutive Network for Human Pose Estimation.

Multi-frame human pose estimation in complicated situations is challenging. Although state-of-the-art human joints detectors have demonstrated remarkable results for static images, their performances come short when we apply these models to video sequences. Prevalent shortcomings include the failure to handle motion blur, video defocus, or pose occlusions, arising from the inability in capturing the temporal dependency among video frames. On the other hand, directly employing conventional recurrent neural networks incurs empirical difficulties in modeling spatial contexts, especially for dealing with pose occlusions. In this paper, we propose a novel multi-frame human pose estimation framework, leveraging abundant temporal cues between video frames to facilitate keypoint detection. Three modular components are designed in our framework. A Pose Temporal Merger encodes keypoint spatiotemporal context to generate effective searching scopes while a Pose Residual Fusion module computes weighted pose residuals in dual directions. These are then processed via our Pose Correction Network for efficient refining of pose estimations. Our method ranks No.1 in the Multi-frame Person Pose Estimation Challenge on the large-scale benchmark datasets PoseTrack2017 and PoseTrack2018. We have released our code, hoping to inspire future research.

Visual Results

On PoseTrack

Comparison with SOTA method

Experiments

Results on PoseTrack 2017 validation set

Method	Head	Shoulder	Elbow	Wrist	Hip	Knee	Ankle	Mean
PoseFlow	66.7	73.3	68.3	61.1	67.5	67.0	61.3	66.5
JointFlow	-	-	-	-	-	-	-	69.3
FastPose	80.0	80.3	69.5	59.1	71.4	67.5	59.4	70.3
SimpleBaseline(2018 ECCV)	81.7	83.4	80.0	72.4	75.3	74.8	67.1	76.7
STEmbedding	83.8	81.6	77.1	70.0	77.4	74.5	70.8	77.0
HRNet(2019 CVPR)	82.1	83.6	80.4	73.3	75.5	75.3	68.5	77.3
MDPN	85.2	88.8	83.9	77.5	79.0	77.0	71.4	80.7
PoseWarper(2019 NIPS)	81.4	88.3	83.9	78.0	82.4	80.5	73.6	81.2
DCPose	88.0	88.7	84.1	78.4	83.0	81.4	74.2	82.8

Results on PoseTrack 2017 test set(https://posetrack.net/leaderboard.php)

Method	Head	Shoulder	Elbow	Wrist	Hip	Knee	Ankle	Total
PoseFlow	64.9	67.5	65.0	59.0	62.5	62.8	57.9	63.0
JointFlow	-	-	-	53.1	-	-	50.4	63.4
KeyTrack	-	-	-	71.9	-	-	65.0	74.0
DetTrack	-	-	-	69.8	-	-	65.9	74.1
SimpleBaseline	80.1	80.2	76.9	71.5	72.5	72.4	65.7	74.6
HRNet	80.0	80.2	76.9	72.0	73.4	72.5	67.0	74.9
PoseWarper	79.5	84.3	80.1	75.8	77.6	76.8	70.8	77.9
DCPose	84.3	84.9	80.5	76.1	77.9	77.1	71.2	79.2

Results on PoseTrack 2018 validation set

Method	Head	Shoulder	Elbow	Wrist	Hip	Knee	Ankle	Mean
AlphaPose	63.9	78.7	77.4	71.0	73.7	73.0	69.7	71.9
MDPN	75.4	81.2	79.0	74.1	72.4	73.0	69.9	75.0
PoseWarper	79.9	86.3	82.4	77.5	79.8	78.8	73.2	79.7
DCPose	84.0	86.6	82.7	78.0	80.4	79.3	73.8	80.9

Results on PoseTrack 2018 test set

Method	Head	Shoulder	Elbow	Wrist	Hip	Knee	Ankle	Mean
AlphaPose++	-	-	-	66.2	-	-	65.0	67.6
DetTrack	-	-	-	69.8	-	-	67.1	73.5
MDPN	-	-	-	74.5	-	-	69.0	76.4
PoseWarper	78.9	84.4	80.9	76.8	75.6	77.5	71.8	78.0
DCPose	82.8	84.0	80.8	77.2	76.1	77.6	72.3	79.0

Installation & Quick Start

Check docs/installation.md for instructions on how to build DCPose from source.

Citation

@InProceedings{Liu_2021_CVPR,
    author    = {Liu, Zhenguang and Chen, Haoming and Feng, Runyang and Wu, Shuang and Ji, Shouling and Yang, Bailin and Wang, Xun},
    title     = {Deep Dual Consecutive Network for Human Pose Estimation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {525-534}
}