This is a TensorFlow implementation of Koller et al.'s "Deep Hand: How to Train a CNN on 1 Million Hand Images When Your Data is Continuous and Weakly Labelled" (CVPR'16) paper.

The pretrained caffe model is converted using ethereon's Caffe to Tensorflow tool.

Tested with:

• Windows 10 - TensorFlow 1.4.0 - Python 3.5 - GTX 960M
• Ubuntu 14.04 - TensorFlow 1.2.1 - Python 2.7 - Titan X

To evaluate the model on One-Million-Hands dataset's test set:

1. Download and extract this repository and the test data to desired locations.
2. Change code_path and data_path accordingly in evaluation.py script.
3. Download the deephand_model.npy model weights and place it in the deephand folder
4. Run python evaluation.py

Once the evaluation is done you should see:

Accruracy on Test Set: 85.4421

This code is set to use the first GPU of your machine. You can easily change it to use any other GPU/CPU by changing the following line in evaluation.py:

with tf.device("/gpu:0"):

Please cite the Deep Hand paper if you use this code in your research:

@inproceedings{koller16:deephand,
  author = {Oscar Koller and Hermann Ney and Richard Bowden},
  title = {Deep Hand: How to Train a CNN on 1 Million Hand Images When Your Data Is Continuous and Weakly Labelled},
  booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year = {2016}
}