Semantic-Segmentation-CNN-for-driving-dataset

We use deep convolutional neural networks to experiment semantic segmentation task on India Driving Dataset. We use pixel accuracy and Intersection-Over-Union as the evaluation metrics and experiment different kinds of methods, including weighted loss, data augmentation, transfer learning, and U-Net architecture. Weighted loss method achieved 61.07% average IoU and data augmentation method with rotated and cropped images achieved 63.73% IoU. However, it largely increases the training time. Using the transfer learning VGG, we achieve 73.74% average pixel accuracy and 58.45% average IoU on the test set. In addition, we achieve 78.03% average test pixel accuracy and 64.04% average test IoU by the method of U-Net. Finally, we build a new model named Mobile-Deeplabv3+, and we achieved 83.56% average pixel accuracy and 71.35% average IoU on the test set.