Detect Landmarks (rtmlib)

How it fits

Typical backends

Caveats

• I/O contractAn empty or zero-size Image short-circuits to an empty result without invoking the network, and when no person is detected the output is an empty list; the outer cardinality is one entry per detected person and varies per frame.
• Hard constraintA run aborts if the loaded model returns an unexpected keypoint width, guarding against a mismatched checkpoint rather than emitting malformed landmarks.
• Accuracymode trades speed for accuracy across lightweight, balanced, and performance presets, while backend selects the inference runtime; both are read once at startup and are not mutable at runtime.
• Limitationdevice is read once at startup and is not mutable; a CPU/GPU switch requires a redeploy.
• FallbackA device starting with `cuda` silently falls back to CPU when CUDA is unavailable at startup.
• CompatibilityWith no `model` slot bound, the detector and pose ONNX for the configured keypoint set and preset are fetched on first start, so outbound HTTPS to the model host is required when the component initializes; binding the optional `model` bundle (the matching detector `.onnx` and pose `.onnx`) loads both offline with no download.