[Paper Review] Active Terahertz Imaging Dataset for Concealed Object Detection

Dong Liang, Fei Xue+1 more|arXiv (Cornell University)|May 8, 2021

Terahertz technology and applications21 references20 citations

TL;DR

This paper introduces the first publicly available active Terahertz imaging dataset for concealed object detection, featuring 3,157 images with 1,349 labeled objects across 11 classes at 5 mm × 5 mm resolution. RetinaNet achieved the highest mAP (91.46% detection rate, 1.27% false alarm rate), outperforming YOLOv3, YOLOv4, and FRCN-OHEM, especially on small and hard-to-detect objects like ceramic knives and keys.

ABSTRACT

Concealed object detection in Terahertz imaging is an urgent need for public security and counter-terrorism. In this paper, we provide a public dataset for evaluating multi-object detection algorithms in active Terahertz imaging resolution 5 mm by 5 mm. To the best of our knowledge, this is the first public Terahertz imaging dataset prepared to evaluate object detection algorithms. Object detection on this dataset is much more difficult than on those standard public object detection datasets due to its inferior imaging quality. Facing the problem of imbalanced samples in object detection and hard training samples, we evaluate four popular detectors: YOLOv3, YOLOv4, FRCN-OHEM, and RetinaNet on this dataset. Experimental results indicate that the RetinaNet achieves the highest mAP. In addition, we demonstrate that hiding objects in different parts of the human body affect detection accuracy. The dataset is available at https://github.com/LingLIx/THz_Dataset.

Motivation & Objective

• To address the lack of public datasets for evaluating multi-object detection in active Terahertz imaging, particularly under low-quality imaging conditions.
• To enable benchmarking of deep learning detectors on real-world concealed object detection tasks with challenging imaging characteristics.
• To investigate the impact of object position, posture, and material on detection performance in Terahertz imaging.
• To provide a standardized dataset linking computer vision and photoelectric imaging communities for advancing detection under harsh imaging conditions.

Proposed method

• Acquired 3,157 active Terahertz images using a 140 GHz array scanning system with 5 mm × 5 mm resolution.
• Collected data from 10 models (4 male, 6 female) with objects concealed in various body locations (arms, chest, legs, etc.) and orientations (parallel/perpendicular to imaging plane).
• Annotated 1,349 objects across 11 classes using Pascal VOC format, including 289 'unknown' class samples.
• Evaluated four state-of-the-art detectors—YOLOv3, YOLOv4, FRCN-OHEM, and RetinaNet—on the dataset using mAP, detection rate (DR), and false alarm rate (FA).
• Applied Focal Loss in RetinaNet to address class imbalance and hard negative samples, improving performance on rare or small objects.
• Conducted image-level evaluation using DR and FA metrics, and analyzed detection recall by object position and orientation.

Experimental results

Research questions

• RQ1 How does object position and orientation on the human body affect detection performance in active Terahertz imaging?
• RQ2 Which deep learning detector achieves the highest mAP and detection rate on a public dataset with low-quality Terahertz images?
• RQ3 How do class imbalance and small object detection impact the performance of standard object detectors in Terahertz imaging?
• RQ4 Can multi-view or enhanced feature fusion improve detection of objects placed perpendicularly to the imaging plane?
• RQ5 What is the trade-off between detection rate and false alarm rate across different detectors on this dataset?

Key findings

• RetinaNet achieved the highest mean average precision (mAP) and detection rate (91.46%) with a low false alarm rate (1.27%), outperforming YOLOv3, YOLOv4, and FRCN-OHEM.
• The detection rate of RetinaNet was over 90% when the false alarm rate was only 1.27%, indicating strong robustness and reliability.
• YOLOv4 had the lowest false alarm rate (0.63%) but a lower detection rate (84.49%) compared to RetinaNet.
• FRCN-OHEM showed the worst overall performance with a detection rate of 83.86% and a high false alarm rate of 18.99%.
• Object detection performance varied significantly by location: recall was lowest for objects on arms (e.g., 0.5000 for RetinaNet) and higher for legs (0.7000), indicating body geometry affects detectability.
• Objects placed perpendicular to the imaging plane had significantly lower recall (e.g., 0.5577 for RetinaNet) than those parallel to the plane (0.6105), highlighting the need for multi-view detection strategies.