I-YOLO: Improved Progressive Feature Pyramid and Wise-IOU for Object Detection

The YOLO algorithm, as the most typical representation of one-stage object detection methods, is based on deep neural networks for object recognition and localization, achieving real-time performance suitable for deployment in detection systems. YOLOV7 outperforms the previous YOLO series in accuracy and speed. However, its detection accuracy is still not very impressive. To address this problem, we propose an improved model (I-YOLO) based on YOLOV7. First, we propose a progressive feature pyramid network with a distillation module, which improves the efficiency of the model and at the same time can reduce the semantic gaps between non-adjacent layers, and further utilizes adaptive spatial fusion operations to alleviate the target information conflict problem when fusing features across layers. Second, we introduce the Wise-IoU loss function to optimize the model and improve the accuracy of the algorithm. Furthermore, we train I-YOLO from scratch only on the 2007 and 2012 datasets of PASCAL VOC without using any other datasets or pre-trained weights. Experimental results show that competitive results are achieved: 55.9% AP (74.7% AP50), and I-YOLO improves by about 2.5% over the baseline model (YOLOV7).