Application of Spatial transformer network to a hand alignment task.
All results refer to the dataset avaiable here.
A full description of this project is provided in ImageAnalysisArticle.pdf
If you want to create your own dataset run acquisition.cpp source file on Windows OS and arrange the dataset folder like we did. The STNtest script includes the code to test the network.
The file acquisition.cpp has the following dependecies:
and obviously a depth camera like Senz3d or DepthSense.
An example of depth frames (8 bit PNG encoding) acquisition is shown in the following gif:
| Package name | Version |
|---|---|
| Python | 2.7 |
| Numpy | 1.14.3 or higher |
| OpenCV | 3.4.1 or higher |
| Tensorflow | 1.7.0 or higher |
| Keras | 2.1.6 or higher |
| matplotlib | 2.2.2 or higher |
| PIL | 3.4.2 or higher |
and SpatialTransformer.py [2]
The notebook STNtest.ipynb file is provided for a user friendly mode to test the application. Pretrained model weights can be loaded from weights_last.h5.
We recommend to use STN_train.py in a real use context to train the network, just type:
python STN_train.py
- [1] https://arxiv.org/abs/1506.02025
- [2] https://github.com/oarriaga/spatial_transformer_networks
This project is licensed under the MIT License - see the LICENSE.md file for details
- Niccolò Bellaccini, nicco.b.93@gmail.com, https://github.com/kinik93
- Tommaso Aldinucci, tommaso.aldinucci@icloud.com, https://github.com/tom1092