Encouraging Success: The Reward-Driven Approach to Object Detection
Aleksandra M. Cvetkovic, Veljko D. Papic · 2023
To date, YOLO algorithms have achieved great results in general-purpose object detection tasks. However, tasks such as screen time estimation where object from a given class can appear at most once per image can cause problems for this network and produce inadequate results. To solve this problem and improve the performance of YOLO networks on screen time estimation tasks, this study proposes an improved version of loss function and a new type of filtering on YOLOv5 framework. This new type of filtering ensures that each class has at most one object present in each image. Improved loss function focuses on the fact that YOLO networks tend to treat many decent predictions as false positives. We mitigate this problem by either ignoring these predictions when calculating loss or even rewarding the network for finding them. Experimental results show a 2.5% increase in precision when detecting characters in scenes from “Dr House” and a 3.4% increase of recall when detecting letters in American sign language.