Enhancing road segmentation model for Asphalt edge detection (This project was done as a part of my internship at Avular Innovations,Eindhoven,Netherlands)
This project investigates different out-of-distribution generalization strategies primarily using data augmentation to implement a binary segmentation model. The model aims to detect the road edge while ensuring consistent segmentation performance in different weather conditions, camera properties, and image perspectives. Out-of-distribution (OOD) generalization denotes the capability of a model to effectively and accurately pre- dict data that deviates from the distribution encountered during its training phase. Achieving adeptness in OOD generalization is essential for ensuring the model’s reliability and effectiveness in real-world deployment scenarios. Data augmentation was adopted and tested through three different experiments. The first experiment analyzes how our model architecture generalizes to OOD data during testing with the help of data augmentation. The second experiment studies the effect of various data augmentation transforms on the road edge segmentation performance. The third experiment explores the effectiveness of employing transfer learning for enhancing road edge segmentation. Our findings show that data augmentation enhances the model’s performance by increasing its robustness to various operating conditions, while also promoting generalization to out-of-distribution data. For more information about the implementation, please refer to the report here
