For an image segmentation model (DeepLabV3), six XAI methods – LIME, Grad-CAM, Seg-Grad-CAM, Saliency, XRAI, and Integrated Gradients – were implemented using established libraries (Captum and Google's Saliency Library) and evaluated using five different Quantus metrics, namely Max-Sensitivity, IROF, Pointing Game, Focus, and Effective Complexity. Images are taken from autonomous driving scenarios. The empirical results show that no single explanatory method outperforms the others across all evaluation criteria. Perturbation-based methods like LIME offer robust explanations, especially for complex objects like people. In contrast, they may lack functional clarity for simpler, uniform classes such as cars. Conversely, the gradient-based methods Grad-CAM and Seg-Grad-CAM provide high attribution relevance and strong localisation for uniform classes but show lower robustness and spatial specificity. XRAI and IG exhibit a balanced performance, while Saliency demonstrates persistent underperformance, attributable to instability and a lack of interpretability. XAI methods’ effectiveness depends on class and context. Method selection needs to be adapted to align with application domain and the target object.