Self-Driving Car Engineer Nanodegree Program Term3 Project 2
My code utilized a Fully Convolutional Network (FCN) to label the pixels of road in images. The FCN architecture is as below. Skip layer from VGG layer 3, 4 and 7 are combined with upsampled 1x1 convolutional layers.
Hyperparameters are chose with balance with accuracy and efficiency. L2 regularizer(1e-3) are used for all layers. Keep_prob is set to 0.5, learning rate = 0.0001. batch_siz =5. With epochs =1, loss = 0.235. When epochs =10, loss = 0.108. Finally we use epochs =50, loss = 0.01~0.02.
All the labeled plots are in plots folder. Majority of them are correctly labeled with road vs non-road.
Here are some examples: