Detect Document Layout (DocLayout YOLO)

How it fits

Typical backends

Caveats

• I/O contractExactly one arm is emitted, fixed at deploy by the connected consumer: the [BoundingBox] arm carries the layout class id, while the [Rectangle<Double>] arm carries geometry only and no class.
• I/O contractA region's class is a numeric class id; the id-to-label map ships with the mounted DocLayout-YOLO weights file and is exposed through the component metadata, not embedded per box.
• Parameter interactionconfidence drops regions scoring below it; lower it to recover faint regions at the cost of false positives. img_size is the square inference resolution and should suit the checkpoint (1024 for DocStructBench).
• FallbackA zero-size input Image short-circuits to an empty list without running the model.
• Fallbackdevice set to GPU on a host without CUDA falls back to CPU, where per-page latency climbs sharply.
• LimitationOne page per call; multi-page PDFs and tiled captures must be split upstream.
• LatencyThe mounted weights file loads once at startup; first-call latency on a fresh container is dominated by model load, not inference.