Semantic Segmentation for Traffic Scene Understanding

General:
This repo contains a comparison between 3 encoder-decoder networks (FCNN, SegNet and UNet) for image segmentation using traffic scenes from the CamVid dataset. The challenge lies in categorizing each pixel within a traffic scene image to one of 32 categories. The encoder is pre-trained using VGG16 weights. Training and test metrics such as binary cross entropy loss, IoU and pixel accuracy are computed and averaged across an entire epoch.

Fig.1 - Segmentation demo.

Training:
To train and replicate results open up terminal and cd into the semantic-segmentation-traffic-scenes folder using : cd semantic-segmentation-traffic-scenes/, then call the respective modelname_train.py from terminal, for example as followed: python unet_train.py. 3 models are supported and can be trained by calling modelname_train.py, that is fcnn_train.py , segnet_train.py and unet_train.py. Periodical updates regarding training and test are printed in console.

Results:
To obtain evaluation results and corresponding figures such as loss, IoU, pixel accuarcy and predictions on images from the test set, run python plots.py from terminal. Figures will be saved in the /figs folder and should look as followed:

Fig.2 - Test and training accuracy as a function of epochs.

Fig.3 - Model comparison: Test and training intersection over unions as a function of epochs.

Fig.4 - Model comparison: Test and training loss as a function of epochs.