Performance Evaluation of CNN Based Pedestrian and Cyclist Detectors On Degraded Images

Abstract

This paper evaluates the effects of input image degradation on performance of image object detectors. The purpose of the evaluation is to determine usability of the detectors trained on original images in adverse conditions. SSD and Faster R-CNN based pedestrian and cyclist detector performance with images degraded with motion blur, out-of-focus blur, and JPEG compression artefacts, most commonly occurring in mobile or static traffic systems. An experiment was designed to assess the effect of degradations on detection precision and cross class confusion. The paper describes the two datasets created for this evaluation, evaluation of a number of detectors on increasingly more degraded images, comparison of their performance, and assessment of their tolerance to different types of image degradation as well as a discussion of the results.This paper evaluates the effects of input image degradation on performance of image object detectors. The purpose of the evaluation is to determine usability of the detectors trained on original images in adverse conditions. SSD and Faster R-CNN based pedestrian and cyclist detector performance with images degraded with motion blur, out-of-focus blur, and JPEG compression artefacts, most commonly occurring in mobile or static traffic systems. An experiment was designed to assess the effect of degradations on detection precision and cross class confusion. The paper describes the two datasets created for this evaluation, evaluation of a number of detectors on increasingly more degraded images, comparison of their performance, and assessment of their tolerance to different types of image degradation as well as a discussion of the results.