Pose & Landmark Evaluation (Auslan exploration)

This is exploratory tooling that asks a single, prior question: is pose and landmark extraction on Auslan video good enough to build anything on at all? Rather than attempting sign recognition, it treats the extraction step itself as the object of study — inspecting how reliably a pipeline recovers body, hand and face landmarks from multimodal footage before any downstream modelling is contemplated. The scope is deliberately narrow: it measures extraction quality and nothing else. It does not classify signs, and by design it does not process or store personal video.\n\n## Why it matters\n\nSign-language research often rushes to recognition on top of landmark tracks whose reliability was never characterised. If extraction quietly degrades on the fast, occluded, two-handed articulation that Auslan actually uses, every result built above it inherits that weakness without anyone noticing. Making extraction quality legible up front means a decision to proceed — or not — can rest on evidence rather than optimism.\n\nThis work sits inside the wider exploratory Auslan direction, and it is early exploration, not an accessibility product. It is not usable as one and is not presented as one. Community governance comes first: the direction is gated on Deaf-community leadership, consent and governance, and the licence is not yet confirmed precisely because those governance decisions have not been made. The evaluation exists so that any later step is taken with eyes open, under community authority.