Hierarchical loss functions for enhanced object detection in hierarchically structured datasets

Object detection in datasets with hierarchical relationships requires advanced methods that surpass traditional classification algorithms. Conventional models using cross-entropy loss across a flattened set of labels treat all errors equally, neglecting critical hierarchical distinctions. To address this, we propose hierarchical loss functions leveraging the structured relationships in defense-oriented ontologies. By weighting penalties based on hierarchical distance, our approach emphasizes operationally significant distinctions. Our methodology also integrates a combination of deep metric representation learning and state of the art object detection techniques to better capture the relationships between different vehicle types. A multiscale feature extraction mechanism is integrated to capture both fine details and broader context, critical for differentiating similar vehicles under challenging conditions, such as camouflage or occlusion. This paper highlights the effectiveness of hierarchical loss functions for defense vehicle detection, demonstrating enhanced accuracy, error sensitivity, and operational utility compared to traditional methods.