A Comparative Analysis of IoU-Based Losses with YOLOv8: Fruit Tree Detection from UAV Images

For this study, We collected UAV-captured images of mango trees from a mango orchard field, and a standard YOLO annotation format dataset was created. The fined-tuned YOLOv8s lightweight model is trained to detect trees with different IoU-based loss functions for the bounding box regression over the image dataset. The model is trained using 1972 images and validated using 494 images. Later, the key indicators, precision, recall, mAP, Training and Validation losses of the bounding box, Training Time (Hrs), speed (PyTorch ms), and epoch count are calculated. Finally, the conclusion was drawn based on the data comparing the different loss functions GIoU, DIoU, CIoU, EIoU, and SIoU, results are encouraging. CIoU exhibited the best overall optimal performance for UAV images tree detection, achieving the lowest validation (0.199) loss and high accuracy with an effective precision (0.934), recall (0.944), and mAP@50 (0.981). However, convergence requires 155 epochs showing effective optimization but with a longer training duration. GIoU and SIoU also performed well, converging at 127 epochs. SIoU, though efficient and precise, has a higher validation loss (0.216) and slightly lower recall (0.925) compared to CIoU. Overall, this paper highlights the necessity of choosing the appropriate loss function based on specific agricultural image processing requirements.