Convolutional Neural Networks (CNNs) have been dominant in segmentation tasks for some time already. Along with improvements to CNNs, different and novel loss functions have been proposed in combination with different layer architectures to create CNNs that are more powerful than ever before. We applied one of these newer architectures, DeepLabv3 to semantic segmentation for road scenes based on the KITTI-360 dataset, with a real-life application of Autonomous Driving. Furthermore, we also take advantage of transfer learning on our dataset based on DeepLabv3’s original training set, COCO 2017. We analyze the results obtained from training different classes and analyze how the network performs using pretrained weights. The point of transfer learning on this segmentation is to understand how well the network performs when given a new cost surface to minimize.
For more, please read our writeup.
note: model ('final_model.pth') must be extracted with 7zip
Additionally, the unzipped file is also on Box for easier download and management.