Detect objects using HuggingFace models with automatic backend selection

Use This When

• Building detection pipelines that need to swap models without changing downstream components
• Leveraging pre-trained HuggingFace models like DETR, YOLOv8, or RT-DETR for quick prototyping
• Comparing different detection architectures while maintaining consistent output format
• Deploying detection systems where model choice depends on accuracy/speed requirements

What It Does

• Auto-detects optimal backend based on model repository tags and configuration
• Supports Transformers pipeline, PyTorch models, ONNX Runtime, and Ultralytics YOLO
• Returns standardized bounding boxes with class IDs and confidence scores
• Optionally filters returned detections by class ID
• Handles model loading, preprocessing, inference, and postprocessing transparently

Works Best With

• Video inputs → this component → track-objects-yolo for persistent tracking across frames
• Detections → segment-image-hf or detect-landmarks for detailed region analysis
• Integration with visualize or check-object-distance for spatial monitoring
• Feeding detect-zone-transition or evaluate-expression for rule-based alerting

Caveats

• Class ID mapping varies by model; verify downstream components use correct label mappings
• ONNX and YOLO backends require GPU for real-time performance on high resolutions
• Pipeline backend adds convenience but slightly more overhead than direct model inference
• Threshold filtering happens post-NMS; adjust threshold based on precision/recall needs