---

name: object-detection-workflows description: "Still-image object detection workflows for COCO and custom ImageAI YOLO models." disable-model-invocation: true metadata: disco-role: operating license: MIT

Object Detection Workflows

Use this sub-skill when the task is still-image object detection with ImageAI:

• COCO pretrained image detection with ObjectDetection and RetinaNet, YOLOv3, or TinyYOLOv3 weights.
• Custom image detection with CustomObjectDetection, YOLOv3 or TinyYOLOv3 weights, and the matching detection config JSON.
• Detection return contracts for saved files versus arrays.
• Object crop extraction as saved paths or in-memory arrays.
• COCO CustomObjects filters, custom-model label filters, confidence thresholds, NMS/objectness thresholds, and display flags.

Do not use this sub-skill for video files, cameras, live streams, callbacks, or training/data conversion. Route video and camera work to sibling sub-skill video-detection-workflows. Route custom model training, dataset validation, and Pascal VOC-to-YOLO conversion to sibling sub-skill custom-training-and-data.

What to read

• Read references/api-reference.md before writing code that calls ImageAI detection APIs. It records verified signatures, model-type mapping, parameter spelling, and return shapes.
• Read references/workflows.md for task recipes: COCO detection, custom detection, filters, file/array output choices, extraction, and helper-script examples.
• Read references/coco-object-classes.md before using CustomObjects so labels match the loaded COCO model family.
• Read references/troubleshooting.md when a model file, JSON config, output path, extraction directory, object filter, or threshold behaves unexpectedly.

Safe helper

The bundled helper scripts/detect_image.py wraps the still-image detection APIs without downloading weights or assuming the current working directory. Use it when a caller wants a repeatable command-line entry point:

python scripts/detect_image.py --help

The helper supports --mode coco|custom, COCO/custom model-type validation, explicit model/config paths, file versus array outputs, extraction, threshold/display flags, COCO/custom object filters, CPU forcing, and JSON summaries.

Operating reminders

• ImageAI 3.x expects PyTorch .pt or .pth weights. TensorFlow-era .h5 weights are intentionally rejected.
• output_type="file" returns detections and writes an annotated image only when output_image_path is supplied. output_type="array" returns a rendered OpenCV/Numpy image array plus detections.
• Extraction with file output writes a sibling extraction directory based on the output image basename plus -extracted; pre-existing directories cause failure.
• Detection dictionaries contain name, percentage_probability, and box_points with [x1, y1, x2, y2] integer coordinates.
• For COCO filters, call detector.CustomObjects(...) with labels from the loaded class list and spaces converted to underscores, for example traffic_light=True or cell_phone=True.
• Current source uses detectObjectsFromImage(custom_objects=...); older examples may mention detectCustomObjectsFromImage, but that method is not part of the verified current still-image API.