Deep Ensembles for Semantic Segmentation on Road Detection

Semantic segmentation is a significant technique that can provide valuable insights into the context of driving scenes. This work discusses several mechanisms: data augmentation, transfer learning, transposed convolutions and focal loss function for improving the performance of neural networks for image segmentation. Experiments on two traditional model architectures – U-net and MobileUNetV2 – are conducted and the results are evaluated in terms of – Intersection-over-Union (IoU) and F-score. The KITTI Road dataset is utilised for training and testing the algorithms on road segmentation. More specifically, data augmentation and the task-specific focal loss provide the highest improvement of 6.68% and 5.23%, respectively. To further enhance segmentation performance, an ensemble scheme is adopted where several models are executed simultaneously and their outputs are fused together to derive the final prediction. Such a design can reduce incorrect predictions of individual models and produce more precise segmentation masks.