hodm,minghuam

Download resource data(models and demo data) from google drive.

    sh download_resources.sh

To train the CNN models, you need three kinds of data: raw images, hand masks and object/hand heatmaps. See examples in $root/raw_data.

Raw images: capture about 1000 images while performing different actions. For better hand segmentation, capture data from different subjects.
Hand masks: you can manually annoate hand regions or use background subtraction to generate hand masks for the raw images.
Heatmaps: the object/hand heatmap contains probability map for object/hand locations encoded in three channels: blue for object of interest, green for left hand, red for right hand. $root/annotation contains a few tools to label object and hands locations and generate heatmaps.

To train the hand segmentation network, go to $root/segmentation_training.

Get Caffe from here and modify the caffe_root in config.py.
If the Caffe is built successfully and the Caffe root path is correct, you can run the demo and see results like the following:
```
 python demo.py
```
Run the following command to resize raw training data and generate training data source text file: training_data.txt.
```
 python prepare_training_data.py
```
Train the hand segmentation CNN. The trained model files are saved in $root/segmentation_training/model.
```
 sh train.sh
```

Set caffe_root in config.py to the same Caffe version. Set hand_model to the path of the trained hand segmentation model.
If the configuration is correct, you can run the demo for hand and object detection:
```
 python demo.py
```
Run the following command to resize raw training data and generate training data source text file: training_data.txt.
```
 python prepare_training_data.py
```
Train the hand/object detection CNN. The trained model files are saved in $root/detection_training/model.
```
 sh train.sh
```

minghuam / hodm Goto Github PK