chainer-pose-hg

Chainer implementation of Stacked Hourglass Networks for Human Pose Estimation

Requirement

• Python 2.7
• Chainer

Prepare data

• LSP dataset
• LSP Extended dataset

Place all images in data/LSP/images

Start training

python src/train.py

Check the options by python src/train.py --help and modify the training settings.

Get predictions

python src/test.py

Default setting is for predicting LSP test set(1000 images).

Also, you can predict any image by specifying image name in an csv file, set --test_csv_fn and --img_dir

Pre-train model

Will be uploaded soon

Note

Current implementation uses 2 stacks of hourglass and 1 residual modules at each location

refer to the '-nStack' and '-nModules' in opts.lua of pose-hg-train

Reference

• pose-hg-train
• http://seiya-kumada.blogspot.tw/2016/03/fully-convolutional-networks-chainer.html
• DeepPose