PAGF/DET-COMPASS

Superpowering Open-Vocabulary Object Detectors for X-ray Vision

ICCV 2025

Pablo Garcia-Fernandez, Lorenzo Vaquero, Mingxuan Liu, Feng Xue, Daniel Cores, Nicu Sebe, Manuel Mucientes, Elisa Ricci

DET-COMPASS

This is the official repository of Superpowering Open-Vocabulary Object Detectors for X-ray Vision (ICCV'25)

Dataset Summary

Object detection in security X-ray scans has advanced significantly in recent years. However, evaluating Open-vocabulary Object Detectors (OvOD) detectors in this modality remains challenging due to the limited number of annotated object categories in existing X-ray benchmarks. This limitation severely constrains the comprehensive evaluation of OvOD methods, which require a broad and diverse category set to assess generalization to unseen object semantics. To address this gap, we introduce DET-COMPASS, a novel benchmark that repurposes the COMPASS-XP classification dataset for object detection through meticulous bounding box annotation. DET-COMPASS comprises 370 distinct object classes, offering an order-of-magnitude increase in vocabulary size over previous X-ray detection benchmarks. Additionally, it provides pixel-aligned RGB images, ensuring precise spatial correspondence across modalities and facilitating the development of multimodal models. Each object is also labeled with a visibility attribute, indicating whether it produces a discernible signature in the X-ray spectrum.

Dataset Structure

DET-COMPASS is provided in two configurations:

• default: The entire dataset is used in a zero-shot setting, where only text labels are available for each object. This configuration is suitable for evaluating open-vocabulary object detectors without access to any visual prototypes.
• raxo: This configuration replicates the setting in the RAXO paper, where a subset of x-ray prototypes for the objects is available. It is intended for experiments that leverage a small number of visual examples per class.

Each sample in the dataset has the following structure:

{
  "rgb_image": PIL.Image,                        // loaded RGB image
  "rgb_width": int,
  "rgb_height": int,
  "xray_images": [                               // list of 5 loaded images, order: Colour, Grey, High, Low, Density
    PIL.Image, PIL.Image, PIL.Image, PIL.Image, PIL.Image
  ],
  "xray_width": int,
  "xray_height": int,
  "objects": {
    "rgb_bbox": [ [float, float, float, float], ... ],   // list of [x_min, y_min, width, height]
    "xray_bbox": [ [float, float, float, float], ... ],  // list of [x_min, y_min, width, height]
    "categories": [ string, ... ],                       // list of text labels
    "rgb_visible": [ bool, ... ],                        // list of booleans
    "xray_visible": [ bool, ... ],                       // list of booleans
    "dangerous": [ bool, ... ],                          // list of booleans
    "IN_ids": [ string, ... ],                           // list of ImageNet IDs (may be empty)
    "WN_ids": [ string, ... ]                            // list of WordNet synset IDs (may be empty)
  }
}

• The xray_images list always follows this order: Colour, Grey, High, Low, Density.
• All bounding boxes are in [x_min, y_min, width, height] format.
• The categories field contains the text label for each object.
• The IN_ids and WN_ids fields provide ImageNet and WordNet synset IDs when available.
• All lists within the objects field are aligned by index: for any index i, the elements rgb_bbox[i], xray_bbox[i], categories[i], rgb_visible[i], xray_visible[i], dangerous[i], IN_ids[i], and WN_ids[i] all correspond to the same object instance in the image.

The repository also includes a classes.csv file listing all classes present in the dataset. This file has the following columns:

• class: Class name (string)
• IN_id: ImageNet class ID (string, may be empty)
• WN_id: WordNet synset ID (string, may be empty)
• dangerous: Whether the class is considered dangerous (boolean)

This structure supports both zero-shot and prototype-based open-vocabulary object detection experiments.

Usage

from datasets import load_dataset
import matplotlib.pyplot as plt
import matplotlib.patches as patches

# 1. Load the dataset
ds = load_dataset("PAGF/DET-COMPASS", name="default", split="test")

# 2. Select a sample
sample = ds[739]

# 3. Get the images
rgb_img = sample["rgb_image"]
xray_colour_img = sample["xray_images"][0]
xray_grey_img = sample["xray_images"][1]
xray_high_img = sample["xray_images"][2]
xray_low_img = sample["xray_images"][3]
xray_density_img = sample["xray_images"][4]

# 4. Get the bounding boxes ([x_min, y_min, width, height])
rgb_bboxes = sample["objects"]["rgb_bbox"]
xray_bboxes = sample["objects"]["xray_bbox"]

# 5. Get the classes
classes = sample["objects"]["categories"]

# 6. Draw bounding boxes
def plot_image(ax, img, bboxes, classes, name=""):
    ax.imshow(img)
    for bbox, cls in zip(bboxes, classes):
        rect = patches.Rectangle(
            (bbox[0], bbox[1]),
            bbox[2], bbox[3],
            linewidth=2, edgecolor='r', facecolor='none'
        )
        ax.add_patch(rect)
        # Draw class name at the top-left corner of the bbox
        ax.text(
            bbox[0], bbox[1] - 5,  # Slightly above the bbox
            cls,
            color='yellow', fontsize=10, weight='bold',
            bbox=dict(facecolor='black', alpha=0.5, edgecolor='none', pad=1)
        )
    ax.set_title(name)
    ax.axis("off")

fig, axs = plt.subplots(3, 2, figsize=(12, 6))

plot_image(ax=axs[0,0], img=rgb_img, bboxes=rgb_bboxes, classes=classes, name="RGB")
plot_image(ax=axs[0,1], img=xray_colour_img, bboxes=xray_bboxes, classes=classes, name="X-ray Colour")
plot_image(ax=axs[1,0], img=xray_grey_img, bboxes=xray_bboxes, classes=classes, name="X-ray Grey")
plot_image(ax=axs[1,1], img=xray_high_img, bboxes=xray_bboxes, classes=classes, name="X-ray High")
plot_image(ax=axs[2,0], img=xray_low_img, bboxes=xray_bboxes, classes=classes, name="X-ray Low")
plot_image(ax=axs[2,1], img=xray_density_img, bboxes=xray_bboxes, classes=classes, name="X-ray Density")

plt.tight_layout()
plt.show()

Citation

If you use DET-COMPASS in your research, please cite:

@inproceedings{garcia2025superpowering,
      title={Superpowering Open-Vocabulary Object Detectors for X-ray Vision}, 
      author={Pablo Garcia{-}Fernandez and Lorenzo Vaquero and Mingxuan Liu and Feng Xue and Daniel Cores and Nicu Sebe and Manuel Mucientes and Elisa Ricci},
      booktitle={Int. Conf. Comput. Vis. ({ICCV})},
      year={2025},
}