Seeing Through the Data: A Statistical Evaluation of Prohibited Item Detection Benchmark Datasets for X-Ray Security Screening
The rapid progress in automatic prohibited object detection within the context of X-ray security screening, driven forward by advances in deep learning, has resulted in the first internationally-recognized, application-focused object detection performance standard (ECAC Common Testing Methodology for Automated Prohibited Item Detection Systems). However, the ever-increasing volume of detection work in this application domain is highly reliant on a limited set of large-scale benchmark detection datasets that are specific to this domain. This study provides a comprehensive quantitative analysis of the underlying distribution of the prohibited item instances in three of the most prevalent X-ray security imagery benchmark and how these correlate against the detection performance of six state-of-the-art object detectors spanning multiple contemporary object detection paradigms. We focus on object size, location and aspect ratio within the image in addition to looking at global properties such as image colour distribution. Our results show a clear correlation between false negative (missed) detections and object size with the distribution of undetected items being statistically smaller in size than those typically found in the corresponding dataset as a whole. For false positive detections, the size distribution of such false alarm instances is shown to differ from the corresponding dataset test distribution in all cases. Furthermore, we observe that one-stage, anchor-free object detectors may be more vulnerable to the detection of heavily occluded or cluttered objects than other approaches whilst the detection of smaller prohibited item instances such as bullets remains more challenging than other object types.