Tiny Object Detection:Comparative Study using Single Stage CNN Object Detectors

Abstract

The ability of vehicles to perceive the surrounding environment is enhanced by their capacity to detect and localize the objects around them, be it traffic signs, road signs, other vehicles, landmarks etc., the vehicle movement and varying object dimension poses huge challenge on detection especially smaller dimension objects from far distance. In the proposed work such objects are detected at far distance using different single stage CNN detectors. A comparative study was done on three models such as SSD300 and its variants, Tiny-YOLO V3 and RetinaNet. The networks were trained using 7082 sample images and tested on 6500 images. Detection was tested on images with varying scale and rotation in realtime with single class object size of 2-3-inch diameter (20*20 pixels to 40*50 pixels) at 3 to 4 meters of distance. Tiny-YOLO V3 performs comparatively better with 60% accuracy, speed of 0.09 SPF and the model size satisfy the memory constraint.