Overview

We offer a benchmark suite together with an evaluation server, such that authors can upload their results and get a ranking. We offer a dataset that contains more than 25,000 pictures, including 15,403 images for training set, 5,000 images for validation set and 5,000 images for testing set. If you would like to submit your results, please follow the instructions on our submission page.

Note: We only display results with relatively detailed descriptions.

Multi-Human Pose Estimation

Metrics

Following MPII, we use mAP(%) evaluation measure.

Ranking

All teams with successful submissions have a placeholder in the leaderboard, and the results of all teams will be released on 10 June. The winner of the challenge is the team with maximal score of mAP.

Method	mAP	Details	Abbreviation	Submit Time
Baseline	55.8	Details Head Shoulder Elbow Wrist Hip Knee Ankle Total 68.1 69.2 61.9 58.3 38.6 51.2 43.5 55.8	Abbreviation Contributors Description "Generative Partition Networks for Multi-Person Pose Estimation”, Xuecheng Nie, Jiashi Feng, Junliang Xing, Shuicheng Yan; arXiv:1705.07422 Generative Partition Network (GPN) solves multi-person pose estimation problem by simultaneously 1) modeling person detection for joint partition as a generative process inferred from all joint candidates and 2) performing local inference for obtaining joint categorizations and associations conditions on the generated person detections. This result is generated by pre-trained model on MPII Multi-Person dataset without finetuning on MHP dataset.	2018-04-12 11:00:00
JDAI-Human	72.2	Details Head Shoulder Elbow Wrist Hip Knee Ankle Total 85.0 79.2 76.0 75.3 59.2 68.3 62.3 72.2	Abbreviation Contributors Description JDAI-Human: Qian Bao, Wu Liu, Tao Wu, Tao Mei.(JD AI Research) We used Faster R-CNN for multi-person detection and fine-tuned stacked hourglass on MHP training dataset. An 8-stacked hourglass network with data augmentation was used as the basic network and the refinement strategy was added for hard keypoints estimation. [1]RMPE: Regional Multi-person Pose Estimation, Fang H S, Xie S, Tai Y W, et al, ICCV2017. [2] Cascaded Pyramid Network for Multi-Person Pose Estimation, Chen Yilun, Wang Zhicheng, Peng Yuxiang, et al, CVPR 2018.	2018-06-10 12:21:00
OSU-Human	59.2	Details Head Shoulder Elbow Wrist Hip Knee Ankle Total 73.2 67.0 62.8 63.9 45.2 55.3 47.2 59.2	Abbreviation Contributors Description OSU-Human: Cheng Zhang(Dept. of Computer Science & Engineering, The Ohio State University) We use YOLOv3 for detection , then use pyramid residual modules for human keypoints estimation.	2018-05-30 09:31:00
RNG	57.8	Details Head Shoulder Elbow Wrist Hip Knee Ankle Total 29.1 71.9 71.0 67.2 44.1 63.0 58.3 57.8	Abbreviation Contributors Description RNG: Yong Qian(Nanjing University) My approach is based on openpose and also improves it with some tricks.	2018-05-31 22:57:00
MJDG	69.9	Details Head Shoulder Elbow Wrist Hip Knee Ankle Total 85.8 78.5 74.2 73.9 54.1 64.0 58.7 69.9	Abbreviation Contributors Description MJDG (Multi-person Joint Detection and Grouping): Sheng Jin (Tsinghua University), Wentao Liu(Tsinghua University) Our method is a pure bottom-up pose estimation method based on “Associative Embedding: End-to-End Learning for Joint Detection and Grouping”. We use 4 stage stack-hourglass network to extract both high-level and low-level features, and predict 1D “tag” to identify its group assignment. The tag is trained to distinguish different people, using pull&push losses. We perform bottom-up pose detection and greedy grouping to obtain the final results. Multi-scale testing (2, 1.5, 1.25, 1, 0.75, 0.5) is used. Unlike previous works, no single person pose estimation model is used to refine the result. Only train-val data from MHP Dataset is used.	2018-06-10 19:25:00