Detect Landmarks (HuggingFace)

How it fits

Typical backends

Caveats

• Parameter interactionThe inference backend is decided once when the checkpoint loads, by inspecting the resolved repository contents: an exported ONNX graph is preferred when present, otherwise the declared transformer architecture selects the ViTPose or SuperPoint path. Switching model at runtime can change the backend as well as the weights.
• Hard constraintRepositories that ship Python modeling files requiring remote-code execution are rejected at load time; only standard ViTPose, SuperPoint, or ONNX checkpoints are accepted via model.
• I/O contractThe SuperPoint path ignores the [BoundingBox] input entirely and emits a single full-image keypoint list regardless of how many boxes arrive, whereas the ViTPose and ONNX paths process each box independently and preserve input order.
• I/O contractAn empty [BoundingBox] input yields no detections on the ViTPose and ONNX paths; the SuperPoint path still emits its one full-image list. On the flat [Landmark] output arm only the single highest-scoring detection survives, so multi-box inputs collapse to one keypoint list.
• AccuracyKeypoints scoring below threshold are dropped per detection, so a detection can carry fewer than the model's full keypoint set while the per-detection outer count is preserved.
• LatencyChanging model rebuilds the backend on the next inference call, producing a one-call latency spike (and a GPU memory reclaim on CUDA hosts); device is read only at startup and a CPU/GPU switch needs a redeploy.