Detect Document Corners (DocAligner)

How it fits

Typical backends

Caveats

• LimitationSingle-document only; a capture containing several pages must be split upstream before detection.
• Hard constraintAn input smaller than two pixels in either dimension returns all four corners at the origin (a zero-area polygon).
• I/O contractWhen no page is found in the frame the detector emits an empty result, which becomes the same zero-area origin quad as a degenerate frame, so every input still yields exactly one output. A non-empty result that is not four corners is a genuine model-contract violation and fails fast rather than being reshaped into a fabricated full-frame quad.
• OrderingCorner order follows the model and is not guaranteed to be canonical top-left to bottom-left; consumers that rely on corner identity must reorder.
• Parameter interactionpad_size adds a border before inference and is subtracted back afterwards, with corners clipped into frame; very small padding can clip corners that sit near the edge.
• State lifetimeThe model selected by model_cfg is loaded once at startup, so first-call latency on a fresh container is dominated by weight load rather than inference.