Novel Deep Learning Based Stacking Models for Multiclass Object Recognition in Drone Images

Object detection models might not be as accurate as they could be when trained on particular datasets. Learning from the predictions of multiple object detection models can help improve their detection accuracy. The learning of predictions to create a new set of predictions is called meta-learning, and the method to metalearn is Stacking Generalization, or just stacking. This paper proposed Stacking models to improve several multiclass object detectors trained the Visdrone2019-DET dataset and find the most accurate model. Four models were trained -YOLOv5x, YOLOv7x, YOLOv8x, and EVA02L. The models trained on one variation of the dataset. The meta-models trained from the object detection predictions were XGBoost and Multilayer Perceptron (MLP). Nearly all the Stacking models using either XGBoost or MLP meta-models improved the mAP of the object detection models trained on VisDrone-split. The most improved Stacking models in both sets used MLP meta-models. The most accurate model was YOLOv8x + EVA02L MLP on VisDrone-split with an mAP50 of 0.609, with an mAP of 0.389, and whose mAP improved from EVA02L by 4.01%.