Advanced Metrics for the Detection Problem on Perspective Transformed Images

Machine Learning-based object detection and tracking are powerful techniques employed in variety of applications. In some cases, the cameras’ point-of-view is rather wide and detected objects are obscured by the perspective (it is often the case in surveillance and technical vision use cases). This paper describes a simple yet effective technique to deal with this problem. The core of the technique is object detection on perspective transformed (warped) images. Such transformation allows to simplify the spatial structure of the image and, therefore, its analysis. In some cases, after the proposed transform a detection problem can be recast as unidimensional. The paper proposes several specialized metrics for measuring the quality of unidimensional detection problem, as well as encompasses several ways to leverage the simplified structure of the image during detection model training. The proposed methods are illustrated by a practical application of positioning wagons on a classification bowl of the sorting freight railway station.