Rebuilting the CMU-OpenPose pose estimatior using Python with OpenCV and Tensorflow.
(The code comments are partly descibed in chinese)
In this work, I used both caffemodel and tensorflow-graph-model, you can download them here, Then place the pretrained models to corresponding directory respectively.
- place
caffe_models\pose\body_25\pose_iter_584000.caffemodel
intopose-estimator-using-caffemodel\model\body_25\
- place
caffe_models\hand\pose_iter_102000.caffemodel
intohand-estimator-using-caffemodel\model\
- place
openpose graph model coco\graph_opt.pb
intopose-estimator-tensorflow\graph_model_coco\
- OpenCV > 3.4.1
- TensorFlow > 1.2.0
- imutils
See the sub-README.md in sub-folder.
- BODY_25 model is faster, more accurate, and it includes foot keypoints.
- COCO requires less memory on GPU (being able to fit into 2GB GPUs with the default settings) and it runs faster on CPU-only mode.
- MPI model is only meant for people requiring the MPI-keypoint structure. It is also slower than BODY_25 and far less accurate.
Body_25 in left, COCO in middle, MPI in right.
See more Output Format details here, and Hand Output Format included as well.
Script in pose-estimator-tensorflow
folder.