Enhancing Semantic Segmentation for Autonomous Driving with Model Stacking and Patch-Based Augmentation

Semantic segmentation is crucial for autonomous driving because it enables vehicles to accurately understand and interpret their surroundings, which is essential for making safe and informed driving decisions. In this paper, we propose two effective strategies to improve the performance of semantic segmentation: model stacking and patch-based data augmentation. Specifically, we stack the predictions of U-Net, U-Net++, and DeepLabV3 and use them as inputs to meta models, aiming to utilize the complementary strengths of these individual models for better prediction. In addition, to address the challenge of detecting small objects, we slice the dataset into smaller segments and apply randomized data augmentation techniques to each segment independently. This approach not only increases the diversity of the training data, but also helps the model to better recognize and predict small objects. We conducted comprehensive experiments on the CamVid dataset and the results demonstrate the effectiveness of our proposed model stacking and patch-based data augmentation methods.