poseaichallenger / mxnet_pose_for_ai_challenger Goto Github PK

Implementation of "Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields" for AI challenger keypoint competition

Jupyter Notebook 99.17% Python 0.83% Shell 0.01%

humanpose aichallenger

mxnet_pose_for_ai_challenger's Introduction

Implementation of "Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields" for AI challenger keypoint competition

Train demo

install cython package

./cython/rebuild.sh

Generate intermediate files

change folder name and json name in pose_io/parse_label.py

path1 = '/data/guest_users/liangdong/liangdong/practice_demo/AIchallenger/keypoint_validation_annotations_20170911.json' 
trainimagepath = '/data/guest_users/liangdong/liangdong/practice_demo/AIchallenger/validation_image/keypoint_validation_images_20170911/'

python pose_io/parse_label.py

Train

python TrainWeight.py

You can download mxnet model and parameters for vgg19 from here

Cite paper Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields

@article{cao2016realtime,
  title={Realtime Multi-Person 2D Pose Estimation using Part Affinity Fields},
  author={Zhe Cao and Tomas Simon and Shih-En Wei and Yaser Sheikh},
  journal={arXiv preprint arXiv:1611.08050},
  year={2016}
  }

Other implementations of Realtime Multi-Person 2D Pose Estimation

Original caffe training model

Original data preparation and demo

Pytorch

keras

mxnet

mxnet_pose_for_ai_challenger's People

Contributors

Stargazers

Watchers

Forkers

gaolingnan guoxpl xzhaogit galaxy-fangfang tonyfaith fendaq upgirlnana luluqie

mxnet_pose_for_ai_challenger's Issues

Training loss

@qqsh0214 How is the training? I realize that it may better to set the batch_size larger.

Hi, I aslo particiate AI Challenger, using the original Caffe implementation of CMU-pose. However the result s are not desirable, the submission obtain a scroe of 41%.
I am wondering what is the performance of your implementation. If you did not get a desirable results either, maybe this kind of method is not fit for this challenger. After all, the evaluation criterion differs from COCO and MPI, which heavily depends on the amount of persons.

mask rcnn

@qqsh0214 It seems to be a good one to work on.
https://github.com/TuSimple/mx-maskrcnn

training error too big.

Training three epochs, and the error is very big.
[6876.2133925787293, 6788.8553634945738, 6788.854836900161]

a problem about computing a person's center

“human_annotations”：若干长度为4的整数型数列，存储人体框的位置。其中前两个参数为人体框左上角点的坐标值，后两个参数为人体框右下角点的坐标值。

"human_annotations"：contains the positions of human bounding boxes. The first two parameters indicate the top left coordinates of the human bounding box, and the last two parameters are the lower right coordinates.

person_center = [box[str(human)][0] + box[str(human)][2]/2, box[str(human)][1] + box[str(human)][3]/2 ]
It is wrong to compute a person's center as this line code, it should be
person_center = [(box[str(human)][0] + box[str(human)][2])/2, (box[str(human)][1] + box[str(human)][3])/2 ]

如何生成ground truth？

您好
我想请问下如何生成PAF的ground truth
在论文中，如何判断一个点是否在limb上，有一个参数是控制
垂直方向的距离的，但是没有明确说是多少。
我在你的这个函数putVecMaps里看到了一个阈值1
如果我理解不错的话dist = np.absolute(ba_x * bc_y - ba_y * bc_x)计算的是
垂直方向的距离。
那么阈值只有1个像素点不是会很小吗？还是我的理解有问题。
谢谢！！

Training dataset

Hi,

I've been trying to download the training dataset from the official website here but the link appears to be completely broken.

Is there a possibility for you to give me access to the dataset? Maybe you have it backed up in a cloud service and you could provide me some download links...

Thanks in advance and hope you can help me.

how to deal the mask

hello,AIChallenger dataset does not give the segmentation,while coco dataset gives it.I want to know how to deal it?