Skip to Content

Object Detection

Image · HuggingFace · Visual

The Object Detection detector runs any HuggingFace object-detection pipeline on image assets. Each detected object becomes a finding with bounding box coordinates, confidence score, and the object label. A severity_map maps specific labels to severity levels for downstream routing.

When to use

  • Sensitive object detection — flag images containing weapons, ID documents, medical devices, or other controlled objects.
  • Workplace safety — detect missing personal protective equipment (PPE) in facility photos or video frames.
  • Brand compliance — locate specific logos or product placements in marketing assets.
  • Data loss prevention — identify screenshots or photos that contain whiteboards, screens, or handwritten notes with potential data exposure.
  • Asset cataloguing — automatically tag images by the objects they contain to support structured search and filtering.

How it works

Classifyre runs the HuggingFace object-detection pipeline on each image asset. Each detected instance above confidence_threshold produces a separate finding with bbox (bounding box coordinates), score, and label in the finding metadata.

Use top_k to limit detections per image to the highest-confidence results. The optional nms_threshold controls non-maximum suppression — lower values remove more overlapping detections. min_box_area discards small-area detections that are likely noise.

Configuration

ParameterTypeRequiredDescriptionDefaultConstraints
modelstringYesHuggingFace hub ID (e.g. 'facebook/detr-resnet-50', 'hustvl/yolos-small') or absolute local directory path.
model_revisionstringNoGit branch, tag, or commit hash when fetching from the HuggingFace hub.null
devicestringNoInference device: 'cpu' (default), 'cuda', 'mps', or a CUDA device string like 'cuda:0'.cpu
confidence_thresholdnumberNoMinimum detection confidence to report an object as a finding (0-1).0.5min 0, max 1
top_kintegerNoKeep only the top-k highest-confidence detections per image.null
nms_thresholdnumberNoIoU threshold for non-maximum suppression. When null the model's default post-processing is used.null
min_box_areaintegerNoMinimum bounding-box area in pixels (width × height). Smaller detections are suppressed.null
severity_maparrayNoOrdered rules mapping detected object labels to severity levels. Labels with no matching rule receive 'info' severity.null

Examples

Detect weapons or sensitive objects

{
  "type": "OBJECT_DETECTION",
  "model": "facebook/detr-resnet-50",
  "confidence_threshold": 0.7,
  "severity_map": [
    { "label": "knife", "severity": "high" },
    { "label": "gun", "severity": "critical" }
  ],
  "top_k": 20
}

PPE compliance check

{
  "type": "OBJECT_DETECTION",
  "model": "keremberke/yolov8m-hard-hat-detection",
  "confidence_threshold": 0.5,
  "nms_threshold": 0.45,
  "min_box_area": 400,
  "severity_map": [
    { "label": "no-hardhat", "severity": "high" }
  ],
  "severity": "info"
}
Last updated on