kevenge / yoloow Goto Github PK

This is the Early Access of the paper "YoloOW: A Spatial Scale Adaptive Real-time Object Detection Neural Network for Open Water Search and Rescue from UAV Aerial Imagery"

In this paper, we proposed a real-time object detection neural network named YoloOW that adapts to different spatial scales for improved object detection performance in open water search and rescue from UAV aerial imagery. Three new model architectures were proposed in the YoloOW model to improve scale adaptability: OaohRep block, DELAN and its enhanced variant E-DELAN, and OaohRepBi-PAN. The YoloOW model underwent quantitative evaluation in accuracy, speed, and model size. It has outperformed all other models with 37.18AP, securing the top position on the leaderboard. Additionally, it achieves a speed of 65.3 FPS, which is faster than any other model with similar accuracy. Moreover, the YoloOW model has only 2.8M more parameters than the lightest model, Yolov7, in model size.

"YoloOW.pt"