Implementations of different variations of U-net - adding deconv layers, dense net variant and including skip connections for the task of generating binary masks. Also, used alpha matting [reference] for post-processing.
Train the model using the implementation files provided and while testing, modify the architecture according to required variant in testing.py.
- The skip connection variant gave best result compared to other variants
- Introducing skip connections led to ~3x improvement in speed