This is a Keras-based Python implementation of DeepMask- a complex deep neural network for learning object segmentation masks. The full article can be found here: Learning to Segment Object Candidates.

This was implemented as a final project for TAU Deep Learning course (2016).

1. Install all requirements, as listed below
2. Download mscoco annotations (see below)
3. Download and convert graph weights with HeplerScripts/CreateVggGraphWeights.py (see below)
4. Create the learning dataset using ExamplesGenerator.py
5. Create a train and test directories with examples to train and test on. Default locations are 'Predictions/train' and same for test (can be configured in EndToEnd.py)
6. Run EndToEnd.py

This was run on Windows 8.1 (64 bit) on a CPU with 8GB RAM. In brackets are the versions I used.

• Python
  • Anaconda for Windows x64 with Python 2.7 (Anaconda version 2.4.1)
• Theano (0.8.0.dev0)
  • conda install mingw libpython
  • git clone https://github.com/Theano/Theano.git
  • python setup.py install
• Keras (0.3.1)
  • NOTICE- if you get an error regarding the 'Graph' object, you are probably not using the correct Keras version
• Open CV (3.1.0)
  • Download installation here
  • Copy cv2.pyd to site-packages dir inside python's lib dir in anaconda
• Coco API (1.0.1)
  • git clone https://github.com/pdollar/coco
  • python setup.py build_ext install

• VGG-D
  • Net was taken from here. Weights should be downloaded locally, and converted to graph format using HelperScripts/CreateVggGraphWeights.py
  • Full article: Very Deep Convolutional Networks for Large-Scale Image Recognition
• MSCOCO
  • Download annotation files here