Can DNN Detectors Compete Against Human Vision in Object Detection Task?

Qiaozhe Li, Jiahui Zhang, Xin Zhao, Kaiqi Huang

Abstract: Object detection is the most fundamental problem in computer vision, which receives unprecedented attention in past decades. Although state-of-the-art detectors achieve promising results on public evaluation datasets, they may still suffer performance degradation problems dealing with some specific challenging factors. In contrast, evidences suggest that human vision system exhibits greater robustness and reliability compared with the deep neural networks. In this paper, we study and compare deep neural networks and human visual system on generic object detection task. The purpose of this paper is to evaluate the capability of DNN detectors benchmarked against human, with a proposed object detection Turing test. Experiments show that there still exists a tremendous performance gap between object detectors and human vision system. DNN detectors encounter more difficulties to effectively detect uncommon, camouflaged or sheltered objects compared with human. It’s hoped that findings of this paper could facilitate studies on object detection. We also envision that human visual capability is expected to become a valuable reference to evaluate DNN detectors in the future.