Insights of semantic segmentation using the DeepLab architecture for autonomous driving

One of the critical tasks of autonomous driving systems is the Perception task (detecting the surroundings), which involves semantic Segmentation. The vital computer vision task of semantic segmentation assigns a "label" to every pixel in the input image. "Semantic segmentation" task consists of partitioning scenes as seen by the Autonomous Vehicle into several communicative slices by categorizing and labelling all image pixel for semantics. This paper gives insights into DeepNet V3 + architecture with ResNet50V2 as the backbone and the other as EfficientNetv2 backbone for feature extraction. The impact of the Squeeze and Excitation module and the Convolutional Block Attention Module is also compared for these architectures for semantic segmentation using the CAMVid data set. All six models are evaluated for Categorical Accuracy and mIoU metrics. The maximum Categorical Accuracy of 97.25 % was achieved in the model ResNet50V2 as the backbone and the Mean IoU of 80.56 %•Feature extraction using DeepNet V3 + architecture with ResNet50V2 and EfficientNetv2 as the backbone.•Insights of using the Squeeze and Excitation and Convolutional Block Attention Module for the DeepNet V3 + architecture.