PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Real-World Prohibited Item Detection: A Large-Scale X-ray Benchmark

Boying Wang, Libo Zhang, Longyin Wen, Xianglong Liu, Yanjun Wu

arXiv:2108.07020v1cs.CV

TL;DR

Inspectors struggle to localize prohibited items hidden in messy objects, and previous datasets often lack comprehensive coverage of such cases. The paper introduces PIDray, a 47,677-image benchmark with 12 categories and item annotations, alongside SDANet. On the hidden test set, SDANet reports 1.5% and 1.3% AP gains over Cascade Mask R-CNN for object detection and instance segmentation, respectively.

  • Problem

    Inspectors struggle to localize prohibited items hidden in messy objects, while prior datasets often lack broad categories, detailed annotations, and cases of deliberate concealment.

    Source: §1 · pp. 1–2

  • Method

    The paper constructs PIDray, a 47,677-image X-ray dataset with 12 prohibited-item categories and bounding-box and mask annotations, and proposes SDANet with dense attention and dependency refinement modules.

    Source: §1 · p. 2

  • Results

    1.5% and 1.3% AP gains over Cascade Mask R-CNN on the hidden test set for object detection and instance segmentation, respectively.

    Source: §1 · p. 2

  • Takeaways & Limitations

    PIDray provides a benchmark aimed at real-world prohibited-item detection, including deliberately hidden cases.

    Source: §1 · p. 2

Abstract

from arXiv · show

Automatic security inspection using computer vision technology is a challenging task in real-world scenarios due to various factors, including intra-class variance, class imbalance, and occlusion. Most of the previous methods rarely solve the cases that the prohibited items are deliberately hidden in messy objects due to the lack of large-scale datasets, restricted their applications in real-world scenarios. Towards real-world prohibited item detection, we collect a large-scale dataset, named as PIDray, which covers various cases in real-world scenarios for prohibited item detection, especially for deliberately hidden items. With an intensive amount of effort, our dataset contains $12$ categories of prohibited items in $47,677$ X-ray images with high-quality annotated segmentation masks and bounding boxes. To the best of our knowledge, it is the largest prohibited items detection dataset to date. Meanwhile, we design the selective dense attention network (SDANet) to construct a strong baseline, which consists of the dense attention module and the dependency refinement module. The dense attention module formed by the spatial and channel-wise dense attentions, is designed to learn the discriminative features to boost the performance. The dependency refinement module is used to exploit the dependencies of multi-scale features. Extensive experiments conducted on the collected PIDray dataset demonstrate that the proposed method performs favorably against the state-of-the-art methods, especially for detecting the deliberately hidden items.

1. Introduction

Inspectors struggle to localize prohibited items hidden in clutter, while earlier benchmarks lack the breadth, annotations, and deliberately hidden cases needed for real-world settings. The paper introduces PIDray and SDANet, reporting gains over Cascade Mask R-CNN on the hidden test set.

1. Introduction

  • Inspectors struggle to localize prohibited items hidden among messy objects, while earlier datasets lack broad categories, detailed annotations, and deliberate-concealment cases.X-ray appearance and object contours differ from natural images, adding challenges for appearance modeling.Source: §1 · pp. 1–2
  • On the hidden test set, SDANet gains 1.5% and 1.3% AP over Cascade Mask R-CNN for object detection and instance segmentation, respectively.The reported gains are for the two subtasks on the deliberately hidden subset.Source: §1 · p. 2
  • SDANet combines spatial and channel-wise dense attention to learn discriminative features with dependency refinement across multi-scale features.The authors report that its attention module is effective for locating deliberately hidden items.Source: §1 · p. 2
  • PIDray provides 47,677 X-ray images across 12 prohibited-item categories, with bounding-box and mask annotations and a test subset for deliberately hidden items.The hidden subset targets items deliberately concealed in messy objects, such as by wrapping wires.Source: §1 · p. 2

2. Related Work

Prior X-ray benchmarks vary in scale, annotation, and complexity, and most do not address deliberately hidden prohibited items. The paper situates its approach among object detectors and attention methods, emphasizing multi-scale feature relations for X-ray detection.

2.1. Prohibited Items Benchmarks

  • Earlier X-ray benchmarks have limitations in scale, category coverage, annotation, and complexity; the paper positions PIDray as a large benchmark with pixel-level annotations and deliberately hidden items.Prior datasets include image-level-only annotation or few images and categories, and few studies target deliberate concealment.Source: §2.1 · pp. 2–3

2.2. Object Detection

  • Object detectors are broadly grouped into two-stage and one-stage methods, with cited examples including Faster R-CNN, FPN, Mask R-CNN, RetinaNet, and anchor-free approaches.The related-work discussion describes FPN as combining low- and high-resolution features, and anchor-free methods as representing objects with key points.Source: §2.2 · p. 3

2.3. Attention Mechanism

  • The paper motivates selective dense attention by combining low-level detail with high-level semantics and learning relations across feature maps at inter-channel and inter-pixel positions.The discussion notes that X-ray images lose texture and appearance information and have overlapping contours.Source: §2.3 · pp. 3–4

3. The PIDray Dataset

PIDray is a large X-ray benchmark of prohibited items, with annotations and test subsets organized to represent different detection difficulties, including deliberately hidden items.

3.1. Data Collection

  • PIDray contains 47,677 images across 12 prohibited-item categories collected using three security inspection machines.The machines came from different manufacturers and produced differences in object and background size and color.Source: §3.1 · p. 4

3.2. Data Annotation

  • Volunteers annotated each instance with a bounding box and segmentation mask, followed by multiple rounds of double-checking.More than 10 volunteers labeled the dataset over two months, with each image taking about three minutes to annotate.Source: §3.2 · pp. 4–5

3.3. Data Statistics

  • The 47,677 images are split into 29,457 training and 18,220 test images, with test cases grouped as easy, hard, or hidden.Easy images contain one prohibited item, hard images contain multiple items, and hidden images contain deliberately concealed items.Source: §3.3 · p. 5

4. Selective Dense Attention Network

SDANet selectively combines features across pyramid levels using channel-wise and spatial attention, then refines the aggregated features to model dependencies.

4.1. Network Architecture

  • SDANet supplements a feature pyramid with selective attention to fuse multi-level features and dependency refinement to enhance them before prediction.The original and enhanced maps are combined and passed to the Region Proposal Network.Source: §4.1 · pp. 5–6, §4 · p. 5

4.2. Selective Channel-wise Attention

  • SCA uses global average pooling and fully connected layers to produce adaptive channel-wise attention weights for feature maps at different pyramid levels.The weighted features are aggregated into an enhanced feature map.Source: §4.2 · p. 6

4.3. Selective Spatial Attention

  • SSA derives spatial context descriptors from average and maximum pooling, then uses spatial attention weights to aggregate feature maps across levels.The weights are computed for each layer and pixel position.Source: §4.3 · p. 6

4.4. Dependency Refinement

  • The dependency refinement module aggregates global context, establishes relationships between channels, and fuses that context into features at all positions.It is applied after channel and spatial attention have aggregated features.Source: §4.4 · p. 6

5. Experiment

Experiments evaluate SDANet on PIDray and natural-image benchmarks, using comparisons and module ablations to assess its performance.

5.1. Implementation Details

  • Comparisons use PIDray training and test sets; SDANet is based on Cascade Mask R-CNN with ResNet-101 and images resized to 500×500.The implementation uses MMDetection, PyTorch, two NVIDIA Tesla V100 cards, and SGD.Source: §5.1 · pp. 6–7

5.3. Overall Evaluation

  • 1.5% and 1.3% AP gains over Cascade Mask R-CNN on the hidden test set accompany SDANet's best performance across PIDray subsets.The evaluation results are reported in Table 3, and Figure 8 compares example outputs with Cascade Mask R-CNN.Source: Table 3 · p. 7, §5.3 · p. 7, Figure 8 · p. 8

5.2. Evaluation Metrics

  • AP averages results across 10 IoU thresholds from 0.50 to 0.95 and 12 categories; AP50 and AP75 use IoU thresholds of 0.50 and 0.75.ARS evaluates small prohibited items, and AR is maximum recall at a fixed number of detections per image, averaged over categories and IoUs.Source: §5.2 · p. 7

5.4. Ablation Study

  • Adding SCA, SSA, and dependency refinement successively improves detection AP by 0.3%, 0.6%, and 0.6%, and segmentation AP by 0.4%, 0.2%, and 0.7%, respectively.On the hidden test set, DR also has reported advantages over SE and CBAM.Source: Table 4 · p. 7, §5.4 · p. 8, Table 5 · p. 8

5.5. Evaluation on General Detection Dataset

  • SDANet gains 0.6 AP on MS COCO [20] and 1.0 AP on PASCAL VOC [7] over Cascade Mask R-CNN.The experiments follow the training and testing pipelines in MMDetection.Source: §5.5 · p. 8, Table 6 · p. 8

6. Conclusion

PIDray is a challenging prohibited-item detection dataset focused on items hidden in other objects, and the paper proposes SDANet to learn the importance of multi-scale feature maps. Experiments on PIDray support the method's superiority.

  • Experiments on PIDray prove the superiority of the proposed method.Source: §6 · p. 8
  • PIDray focuses on prohibited items hidden in other objects and is, to the authors' knowledge, the largest prohibited-item detection dataset so far.Source: §6 · p. 8
  • All PIDray images have instance bounding-box and mask annotations.Source: §6 · p. 8
  • SDANet is proposed to learn the importance of multi-scale feature maps.Source: §6 · p. 8
  • The authors hope PIDray will support a unified evaluation platform and plan to add images and richer annotations for comprehensive evaluation.Source: §6 · p. 8

Related papers

Loading 2108.07020v1…