W-Net

W-Net: A Deep Model for Fully Unsupervised Image Segmentation

Tensorflow Implementation of W-Net

Both the loss functions Soft-N-Cut-Loss and Reconstruction Loss has been Implemented

Batch Processing is enabled

As this model is unsupervised, we will be needing a huge amount of dataset to train properly.

To start the training, run

python soft_n_cut_loss.py

For tensorboard visualizations,

cd checkpoints/logs
tensorboard --logdir=.

Loss Description

Training Procedure

Results

Pipeline of Tasks

• Encoder Decoder Architecture
• Reconstruction Loss
• Soft N Cut Loss
• Batch Processing
• Post Processing
  • Heirarchical Segmentation
  • Conditional Random Fields