Openpose from CMU implemented using Tensorflow. It also provides several variants that have made some changes to the network structure for real-time processing on the CPU.
Original Repo(Caffe) : https://github.com/CMU-Perceptual-Computing-Lab/openpose
Features
[x] CMU's original network architecture and weights.
[] Post processing from network output.
[] Faster network variants using mobilenet, lcnn architecture.
[] ROS Support.
You need dependencies below.
-
python3
-
tensorflow 1.3
-
opencv 3
-
protobuf
| Dataset | Model | Description | Inference Time 1 core cpu |
|---|---|---|---|
| Coco | cmu | CMU's original version. Same network, same weights. | 3.65s / img |
| Coco | dsconv | Same as the cmu version except for the depthwise separable convolution of mobilenet. | 0.44s / img |
| Coco | mobilenet | ||
| Coco | lcnn |
CMU Perceptual Computing Lab has modified Caffe to provide data augmentation.
This includes
-
scale : 0.7 ~ 1.3
-
rotation : -40 ~ 40 degrees
-
flip
-
cropping
See : https://github.com/CMU-Perceptual-Computing-Lab/caffe_train
[1] https://github.com/CMU-Perceptual-Computing-Lab/openpose
[2] Training Codes : https://github.com/ZheC/Realtime_Multi-Person_Pose_Estimation
[3] Custom Caffe by Openpose : https://github.com/CMU-Perceptual-Computing-Lab/caffe_train
[2] Pretrained model : https://github.com/tensorflow/models/blob/master/slim/nets/mobilenet_v1.md
[1] Tensorpack : https://github.com/ppwwyyxx/tensorpack