[Paper Review] Multi-Class 3D Object Detection Within Volumetric 3D Computed Tomography Baggage Security Screening Imagery

Qian Wang, Neelanjan Bhowmik+1 more|arXiv (Cornell University)|Aug 3, 2020

Advanced X-ray and CT Imaging36 references4 citations

TL;DR

This paper proposes a unified 3D RetinaNet framework for multi-class 3D object detection in volumetric CT baggage screening imagery, leveraging data augmentation and backbone network variations to achieve a mean Average Precision (mAP) of 65.3% across five prohibited object classes—bottles, handguns, binoculars, glock frames, and iPods—despite challenges from data scarcity and clutter resemblance.

ABSTRACT

Automatic detection of prohibited objects within passenger baggage is important for aviation security. X-ray Computed Tomography (CT) based 3D imaging is widely used in airports for aviation security screening whilst prior work on automatic prohibited item detection focus primarily on 2D X-ray imagery. These works have proven the possibility of extending deep convolutional neural networks (CNN) based automatic prohibited item detection from 2D X-ray imagery to volumetric 3D CT baggage security screening imagery. However, previous work on 3D object detection in baggage security screening imagery focused on the detection of one specific type of objects (e.g., either {\it bottles} or {\it handguns}). As a result, multiple models are needed if more than one type of prohibited item is required to be detected in practice. In this paper, we consider the detection of multiple object categories of interest using one unified framework. To this end, we formulate a more challenging multi-class 3D object detection problem within 3D CT imagery and propose a viable solution (3D RetinaNet) to tackle this problem. To enhance the performance of detection we investigate a variety of strategies including data augmentation and varying backbone networks. Experimentation carried out to provide both quantitative and qualitative evaluations of the proposed approach to multi-class 3D object detection within 3D CT baggage security screening imagery. Experimental results demonstrate the combination of the 3D RetinaNet and a series of favorable strategies can achieve a mean Average Precision (mAP) of 65.3\% over five object classes (i.e. {\it bottles, handguns, binoculars, glock frames, iPods}). The overall performance is affected by the poor performance on {\it glock frames} and {\it iPods} due to the lack of data and their resemblance with the baggage clutter.

Motivation & Objective

• Address the challenge of detecting multiple prohibited object classes simultaneously in 3D CT baggage imagery, moving beyond single-class detection.
• Develop a unified deep learning framework to reduce the need for multiple specialized models in real-world aviation security screening.
• Improve detection performance in volumetric 3D CT data by exploring data augmentation and backbone network variations.
• Investigate the impact of input resolution scaling and anchor size adaptation on detection accuracy in 3D object detection.
• Evaluate the role of dual-energy (low/high) raw CT data in enhancing detection performance for prohibited items.

Proposed method

• Employ 3D RetinaNet as the primary object detection framework, integrating feature pyramid networks (FPN) for multi-scale detection.
• Use 3D Threat Image Projection (TIP) and data augmentation (3D flipping, rotation) to synthesize training data and mitigate data sparsity.
• Experiment with different 3D ResNet backbones of varying depths to evaluate their impact on detection performance.
• Optimize input resolution by scaling CT volumes with a factor of 3, balancing computational cost and feature representation.
• Adapt anchor sizes across FPN levels to match ground truth bounding box dimensions, improving localization accuracy.
• Evaluate dual-energy CT data (low and high energy) as input channels to assess material contrast benefits for detection.

Experimental results

Research questions

• RQ1 Can a unified deep learning framework effectively detect multiple prohibited object classes in 3D CT baggage imagery, rather than requiring separate models per class?
• RQ2 How do data augmentation techniques such as 3D flipping and rotation affect the generalization and robustness of 3D object detectors in low-data regimes?
• RQ3 What is the optimal input resolution scaling factor and anchor size configuration for 3D object detection in volumetric CT data?
• RQ4 Does combining low- and high-energy raw CT data improve detection performance for prohibited items?
• RQ5 What are the primary causes of failure in detection, particularly for objects like glock frames and iPods, and how do they relate to data scarcity and visual similarity to clutter?

Key findings

• The proposed 3D RetinaNet framework achieved a mean Average Precision (mAP) of 65.3% across five object classes in real volumetric 3D CT baggage imagery.
• The best detection performance was achieved with a 3× input volume scaling factor, as higher or lower resolutions degraded performance despite anchor size adaptation.
• Data augmentation via 3D flipping and rotation significantly improved detection robustness, especially in low-data regimes.
• The detection performance was notably lower for glock frames and iPods due to insufficient training data and visual similarity to baggage clutter, resulting in high false positive rates.
• Combining low- and high-energy raw CT data did not improve detection performance, suggesting that either energy source alone is sufficient for effective detection.
• False positives and overlapping predictions between similar objects (e.g., binoculars and iPods) were frequently observed, likely due to artefacts in synthetic 3D TIP data being mislearned by the model.