In this project, a fully convolutional neural network (FCN) was implemented to perform semantic segmentation of the pixels of road in the images from Kitti Road Dataset.
For this semantic segmentation task I used a pre-trained VGG-16 network (trained on 'imagenet') and adding additional 1x1 layers and skip conections to build a FCN that was used for training. The total number of classes for this task is two as we are performing a binary classification task to segment road vs non-road pixels from the images. Final convolution layer from VGG-16 i.e. layer 7 was input to 1x1 convolution with depth being equal to the desired clases. The other intermediate convolution layers from VGG-16 i.e. layer3 and layer4 are used as skip connection along with 1x1 convolutions and upsampled later to finally give the network output.
HYPERPARAMETERS:
Batch size = 5 Training Epochs = 50 L2 regularizer = 1e-3 learning rate = 0.0009 keep probability = 0.5
The final cross entropy loss after 50 epochs was reduced to 0.026 !
The results on the test set using the save_inference_samples from the helper function:
main.py will check to make sure you are using GPU - if you don't have a GPU on your system, you can use AWS or another cloud computing platform.
Make sure you have the following is installed:
Download the Kitti Road dataset from here.
Extract the dataset in the data folder. This will create the folder data_road with all the training a test images.
Run the following command to run the project:
python main.py
Note If running this in Jupyter Notebook system messages, such as those regarding test status, may appear in the terminal rather than the notebook.