FCN car segmentator

A car segmentation fully convolutional network

Getting Started

Prerequisites

In order to execute this code the following python libraries are needed:

• tensorflow
• ...
• otros

Easy way

Install the necessary packages to get the dataset:

$ apt install wget unzip

Then, inside de repository folder execute

$ data/get_data.sh

Manual way

Alternatively, you can download the dataset from here, and decompress it in data/ig02-cars.

Running the network

The code is distributed in four files:

• fcn.py
• fcn_train.py
• fcn_eval.py
• fcn_input.py

fcn.py

The network architecture is implemented in this file. If you want to change something related to the architecture (layers, loss function, etc) this is the place.

fcn_input.py

The dataset preprocessing happens here, to modify anything related with the input to the network, you can do it here.

fcn_train.py

This is the script that must be executed to train the network.

$ python fcn_train.py

Attention: This script will destroy previous training checkpoints

If you want to watch the progress, run

$ tensorboard --logdir=/tmp/ig02_train

You can change the logdir modifing a line in fcn_train.py, remember to modify accordingly fcn_eval.py as well.

fcn_eval.py

This is the script that evaluates the network being trained using the evaluation set.

$ python fcn_eval.py

Attention: This script will destroy previous evaluating checkpoints

If you want to watch the progress, run

$ tensorboard --logdir=/tmp/ig02_eval

References

• Fully Convolutional Networks for Semantic Segmentation
• Focal Loss for Dense Object Detection
• Convolutional Neural Network - Tensorflow