A Novel Approach for Small-Object Detection based on Fine-tuning and Sliced Inference

Small object detection has a wide range of applications in a variety of fields, including unmanned aerial vehicles (UAVs), surveillance, agriculture, and many more. A significant problem in surveillance applications is the detection of small objects and items that are located far away in the image. As a result, they are difficult to identify with traditional detectors since they are represented in the image by a small number of pixels and insufficient information. Investigation of the state-of-the-art performance of YOLO-based object detection models for the problem of small object detection is motivated by the fact that they are one of the most popular and straightforward object detection models to use. In this research, YOLOv5 and YOLOX models have been used. Aside from that, the consequences of sliced inference as well as the effects of modifying the model for slicing aided inference have been taken care of. The model was trained and evaluated using the VisDrone2019-Det dataset. This dataset presents a challenge in that the majority of the items are small when compared to the image sizes in the dataset. The purpose of this work is to contrast and compare the YOLOv5 and YOLOX models for small item detection. The use of sliced inference has been shown to raise the AP50 score in all tests; however, this effect was more pronounced for the YOLOv5 models when compared to the YOLOX models. A significant improvement was observed for all models when the combined effects of slice fine-tuning and slice inference were considered. A total of 48.8 points was obtained by the YOLOv5-Large model on the VisDrone2019-Det testdev subset, making it the model with the highest AP50 score.