Skip to content

Latest commit

 

History

History
86 lines (79 loc) · 2.85 KB

README.md

File metadata and controls

86 lines (79 loc) · 2.85 KB

Monocular-3d-Hand-Pose-Estimation

Implmentation of Monocular 3d Hand Pose Estimation using ResNet architecture. 😃
The dataset used is FreiHAND dataset which follows the MANO format for the poses. Hence, this model can be easily integrated with SMPL-X or SMPL-H models. 😉
MANO model is used for rendering which was released by Max Plank Institute of Intelligent Systems.

Requirements

  • torch >= 1.6.0
  • torchvision >= 0.7.0
  • trimesh >= 3.7.14
  • pillow >= 7.2.0
  • opencv >= 4.4.0

Steps to Run

  • Download the training and evaluation dataset from here and place it in $(ROOT). Place the testing images in the "evaluation/temp" folder.
    Download the checkpoint file from here and place it in "checkpoints" folder. (Note: The checkpoint file in the link is not fully trained. 😓)
    Final directory structure below 👇 👇
$(ROOT)
  |__ training
            |__rgb
                 |__ 000000.jpg
                 |__ 000001.jpg
                 |__ 000002.jpg
                        ...
   |__ evaluation
            |__rgb
                 |__ 000045.jpg
                 |__ 000046.jpg
                 |__ 000047.jpg
                        ...
            |__ temp
                 |__ 000100.jpg
                 |__ 000101.jpg 
                        ...
   |__ _checkpoints 
            |__ checkpoint_augmented_90.pth
   |__ validation.py
   |__ train.py
   ...
  • (Optional) Change the training parameters to your needs in the train.py. 😇
  • Train the model. 😎
python train.py
  • Validate the model. 😎
python validation.py

Results

Now for the interesting part ❗ ❗ Take a look at the results obtained. 💖

Final Note

The outputs are fairly accurate, but would perform even better if trained for more epochs. The checkpoint file given above is only trained for 90 epochs. 😁

Go ahead..pull it, train it and have fun. And don't forget to ⭐star⭐ the repo, if you like it.😄

🌟 Happiness should be a function without any parameters 🌟

Happy Coding ❗ ❗