YOLO-DynamicNet: A Dynamic Network for Maritime Tiny Object Detection

Maritime object detection is a representative task of tiny object detection within a large field of view. The research on end-to-end object detection networks based on high-resolution images captured by Unmanned Aerial Vehicles (UAVs) further exacerbates the performance degradation of tiny object detection. The excellent YOLOv8 aim to achieve a balance between speed and performance on the benchmark, such as COCO. For the series of YOLO model, the detection of medium and large objects is much excellent, but the detection of tiny objects is given relatively less consideration. To this end, we proposed the YOLO-DynamicNet based on the YOLOv8 to enhance the detection performance of maritime tiny objects. The improved model consists various improvements, such as a tiny object detection layer, the optimization of CSP layer based on dynamic convolution, and the application of dynamic upsampler and dynamic head. The experiments are implemented on the popular SeaDronesSee dataset. The dynamic ability of the network for feature extraction and the improvement of model expressive capacity are demonstrated. YOLO-DynamicNet improves the performance of detecting maritime tiny objects with a relatively small increasing of parameters and computational cost. Compared with the baseline, the performance of the proposed dynamic model is improved by 8.6% in terms of mAP50. Regarding the three typical maritime objects of swimmer, lifesaving appliances, and buoy, the performance is improved by 5.3%, 16.8%, and 18.7% respectively in terms of mAP50.