A Comparative Analysis of Class Imbalance-Aware Loss Functions for Semantic Segmentation of Off-road Scenarios for Autonomous Vehicles

Semantic segmentation in off-road environments is important for autonomous navigation but remains challenging due to severe class imbalance. This imbalance biases deep learning models toward majority classes, resulting in poor recognition of minority classes. This paper presents a comparative analysis of eight loss functions designed for class imbalance, using the Rellis-3D dataset. Our results show that region-based losses, particularly Dice Loss (mIoU - 0.409), achieve the best performance, closely followed by the hard-mining Top-K loss. These findings demonstrate that spatial overlap and difficult examples are more effective for off-road semantic segmentation than conventional frequency-based reweighting strategies.