Performing multi-class semantic image segmentation using a fully convolutional network (FCN-8), a skip architecture that combines semantic information from a deep, coarse layer with appearance information from a shallow, fine layer. This model was proposed by Shelhamer, Long & Darrell in this research paper.
- Original research paper: Fully Convolutional Networks for Semantic Segmentation
- The dataset used is the sample data prepared by Divam Gupta. It can be found in
./dataset
or here. - Pretrained VGG16 weights can be downloaded using Keras. They have been stored in
./pretrained_weights
as well.
Note: to use ./src/predict.py
, first follow the instructions given in ./checkpoints/instructions.md
.
The detailed explanation of the model can be found in the aforementioned paper. The model described in the paper:
Visualization of model returned by ./src/create_model.py
:
Model tested on data in ./dataset/test/
. An accuracy of 86.76% was obtained. Some of the results are as follows: