Human Pose Estimation by Deep Learning

A simple regression based implementation/VGG16 of pose estimation with tensorflow (python) based on JakeRenn's repository. Please read my post for more details about approaches, results, analyses and comprehension of papers: S.-E. Wei, V. Ramakrishna, T. Kanade, and Y. Sheikh. Convolutional pose machines. In CVPR, 2016.

How to Run

The images in data/input/ is not complete. If you just want to run the code with demo images, the codes can run without modification. If you want to train and test with complete images:

• human-pose-estimation-by-deep-learning/data/input/test_imgs: ~4000 test images

• human-pose-estimation-by-deep-learning/data/input/train_imgs: ~60000 train images, use half of them. Each of them are labeled with 15 articulate points.

Please contact with JakeRenn for data (maybe not available currently). And modify the parametaer TAG = "_demo" to TAG = "" in corresponding files like train.py, test.py and draw_point.py.

Train

cd human-pose-estimation-by-deep-learning
python train.py # or: nohup python train.py &

A directory log will be generated to log every training. A directory params_demo will also be generated to store your models. With the log and models, you can use tensorboard for visualisation.

tensorboard --logdir ./log/train_log/ --port=8008# then visit: http://hostname:8008/

Also, you can modify parameters like batch size, number of max training iteration, number of checkpoint iteration in train.py.

Test

python test.py

Then an annos file with predicted position of demo test images should be generated in human-pose-estimation-by-deep-learning/labels/txt/output/test_annos_demo.txt".

Utils

Files in human-pose-estimation-by-deep-learninglabelspython are some utils (in python). For example, you can use draw_point.py to draw the points indicated with annos file to images.

cd human-pose-estimation-by-deep-learninglabelspython
python draw_point.py

Then the images with 15 points should be generated in human-pose-estimation-by-deep-learningdataoutputtrain_imgs_demo (I have generated them for your reference). Modify the parameter PHASE to test will draw test images' points.

For example, the annos file of image 01060538.png is

01060538.png 137,68,136,80,138,105,121,90,107,102,87,114,152,90,168,105,189,118,133,131,130,163,128,201,146,130,149,161,153,199,1

This line means image 01060538.png is labeled with 15 points and their coordinates are (137,68), (136,80), ..., (153,199) (please see the following example output to understand their order). And this line is ended with '1'. By running draw_point.py, the output is

drawing image 01060538.png
coordinate of point # 1 : ( 137 , 68 )
coordinate of point # 2 : ( 136 , 80 )
coordinate of point # 3 : ( 138 , 105 )
coordinate of point # 4 : ( 121 , 90 )
coordinate of point # 5 : ( 107 , 102 )
coordinate of point # 6 : ( 87 , 114 )
coordinate of point # 7 : ( 152 , 90 )
coordinate of point # 8 : ( 168 , 105 )
coordinate of point # 9 : ( 189 , 118 )
coordinate of point # 10 : ( 133 , 131 )
coordinate of point # 11 : ( 130 , 163 )
coordinate of point # 12 : ( 128 , 201 )
coordinate of point # 13 : ( 146 , 130 )
coordinate of point # 14 : ( 149 , 161 )
coordinate of point # 15 : ( 153 , 199 )