Performance Comparison of Convolutional Neural Network Models for Object Detection in Tethered Balloon Imagery

• Débora F. Dos Santos ITA
• André O. Françani ITA
• Marcos R. O. A. Maximo ITA
• Arthur S. C. Ferreira ALTAVE

Resumo

The growing field of remote sensing requires object detection solutions. This work compares the performance of different neural networks for detecting vehicle, person, and boat objects in tethered balloon imagery. Four state-of-the-art architectures were chosen – Faster R-CNN, SSD, Retinanet, and YOLO. For training, the algorithms apply transfer learning, backbone variation, and hyper-parameter tuning. The evaluation utilizes the mean average precision and the inference time as metrics. The YOLOv4 416x model achieved the best trade-off between speed and accuracy with a mAP50 of 90.3%.

Palavras-chave: Training, Transfer learning, Neural networks, Boats, Object detection, Detectors, Real-time systems