AFSDet: Video Small Object Detection Based on Adaptive Focused Slicing

The detection of video small objects is challenging in images due to their limited presence, which makes the information more susceptible to loss. In this paper, we propose an Adaptive Focused Slicing Detection (AFSDet) method inspired by the way humans observe small objects. AFSDet involves first localizing to a region with potential targets and then scrutinizing the small objects within that region. It is a two-stage approach utilizing the proposed model called Deformable Head YOLO Network (DHYNet) for object detection. The first stage performs a coarse-detection of the original image and utilizes clustering and Adaptive Scale-up Slicing methods to obtain patches, leading to a significant reduction in the number of patches compared to previous slicing-based detection methods.The second stage feeds each patches into the fine-detection model and passes the results to the next frame as a reference to assist coarse-detection. Additionally, taking advantage of the slight variations between consecutive video frames, a superimposing process is employed for coarse-detection, which also contributes to reducing the computational overhead. The results show that the proposed method, tested on the VisDrone-MOT dataset, improves accuracy by 9.6% compared to the baseline model, and improves the accuracy by 3.3 % compared to the previous slicing-based small object detection method.