• Stars
    star
    183
  • Rank 210,154 (Top 5 %)
  • Language
    Python
  • License
    Apache License 2.0
  • Created over 5 years ago
  • Updated over 4 years ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

A lite-version hopenet for head pose estimation with PyTorch

Hopenet-lite

A lite-version hopenet for head pose estimation with PyTorch

Note

Hopenet-lite uses unofficial-implement ShuffleNetV2 as backbone network, and now the lastest PyTorch contains official ShuffleNetV2 with various width. If you are seeking for stable performance, please use official implementation and re-train hopenet-lite! I have uploaded the lastest hopenet-lite model with official ShuffleNetV2 from Pytorch torchvision (Please refer to the "Update" part)
'''
import torchvision.models as models
shufflenet = models.shufflenet_v2_x1_0(pretrained=True)
...
https://pytorch.org/docs/stable/torchvision/models.html#classification

Doc.

The project is based on natanielruiz's excellent work named Hopenet.

The link: https://github.com/natanielruiz/deep-head-pose

You can run the network on CPU (i7-8700 six cores) with 35 FPS or GPU (RTX 2070) with 130 FPS

If you used natanielruiz's code in your project, then do not need to change anything except the nueral network you used. At the same time, please refer to the training code of natanielruiz's project for training your own model (if you need)

Here is a trick that affects the accuracy of model: when you get a face's box by any face detectors, cropping a proper face image as the input that may generate better results.

'''
import hopenetlite_v2
net = hopenetlite_v2.HopeNetLite()
saved_state_dict = torch.load('hopenet_lite_6MB.pkl', map_location="cpu")
net.load_state_dict(saved_state_dict, strict=False)
net.eval()
'''

The Pre-trained model in "model" folder, but the model is not very robust to image quality, we will release more robust model in the future.

Thanks for natanielruiz's excellent work again.

Update

Hi, guys, I finally have time to update this project...
I uploaded the lastest hopenet-lite model with official ShuffleNetV2 from Pytorch torchvision, you can use it like this:
'''
import stable_hopenetlite
pos_net = stable_hopenetlite.shufflenet_v2_x1_0()
saved_state_dict = torch.load('model/shuff_epoch_120.pkl', map_location="cpu")
pos_net.load_state_dict(saved_state_dict, strict=False)
pos_net.eval()
'''
The Pre-trained model named "shuff_epoch_120.pkl" in "model" folder. If you think my training is not perfect, you could re-train the model. Just enjoy yourself !

Here are some examples: