Ensemble learning and EigenCAM-based feature analysis for improving the performance and explainability of object detection in drone imagery

Object Detection (OD) from drone images is a long-standing challenge due to multi-scale objects in the images. Several sophisticated and robust OD models have been proposed in the last few decades. This has significantly improved the performance of the OD task. However, these models remain highly opaque. It is challenging for humans to comprehend their outcomes, raising serious concerns about their real-world usability and adoption in mission-critical, high-risk applications. This work investigates the challenge of OD and explainability in drone imagery, aiming to improve the accuracy, robustness, reliability, and trustworthiness of automated OD systems for intelligent surveillance. Our work explores an overview of existing approaches for OD in drone imagery, highlighting their strengths and limitations. The proposed methodology and custom model architecture leverage an integrated pipeline for explainability and ensembling, enabling users to have an improved and better understanding of the OD outcomes. The application of the proposed methodology is demonstrated on the AU-AIR dataset. Significant improvements in object detection accuracy and interpretability are observed compared to existing state-of-the-art methods. The affirmative voting strategy resulted in a 3% increase in mean average precision, demonstrating the potential of ensemble learning to improve the performance of multi-scale OD. The perturbation-based ablation probing of the model with EigenCAM attributes to necessary features reliance on the proposed model with evaluation of XAI for robust, trustworthy, and improved OD outcomes.