Performance Evaluation of YOLOv10 Variants for Deep Learning-Based Automatic Waste Detection

The increasing amount of unmanaged waste poses serious environmental and health risks worldwide. To support automated waste management and enhance waste identification efficiency, this study evaluates six variants of the YOLOv10 object detection architecture—YOLOv10n, YOLOv10s, YOLOv10m, YOLOv10b, YOLOv10l, and YOLOv10x—for automatic waste detection. All models were trained and tested on the Trash Annotations in Context (TACO) dataset, comprising 5,930 annotated objects across 18 waste categories. Each model was trained for 100 epochs with an input resolution of 640×640 pixels using the stochastic gradient descent (SGD) optimizer and a learning rate of 0.01, with the dataset divided into 70% training, 20% validation, and 10% testing subsets. Performance was evaluated using precision, recall, and mean Average Precision (mAP) at IoU thresholds of 0.5 and 0.5:0.95. Experimental results show that YOLOv10l achieved the best overall performance, with precision of 75.3%, recall of 52.5%, [email protected] of 63.9%, and [email protected]:0.95 of 47.3%. Larger models such as YOLOv10x also produced competitive results but with higher computational cost, while lighter variants (YOLOv10n, YOLOv10s, YOLOv10m) exhibited lower detection accuracy. The findings indicate that YOLOv10l offers the best balance between model complexity and detection performance, making it the most suitable variant for waste detection on moderately sized datasets. This research demonstrates the potential of YOLOv10-based architectures to support automated waste management and contribute to environmental sustainability.