Stacked Generalization Models for Improving Multiclass Object Detection in Drone Images

Drones have become increasingly popular due to their many real-world applications such as military, surveillance, agriculture. Automatic detection of objects in drone camera footage poses challenges due to varied angle, heights, occlusion, and illumination. Training a machine learning model on other machine learning model predictions is called meta-learning, and meta-learning from multiple machine learning models is called Stacked Generalization. This research used this technique to propose eight Stacked Generalization models to improve on the three object detection models - Faster R-CNN, RetinaNet, and YOLOv5s. The regression models trained as meta-models were XGBoost and Multilayer Perceptron (MLP). Experiments were performed on the challenging VisDrone image dataset. All the proposed Stacked Generalization models improved from the object detection models they learned from, thereby performing better than the state-of-the-art base models. The best regression model as a meta-model was MLP. The best Stacked Generalization model in terms of accuracy was FRCNN-s + RetinaNet-s + YOLOv5s-s with MLP meta-models, which scored an mAP50 of 0.421. Its mAP was 0.244, which improved 4.72% from YOLOv5s.