Estimate Metric Depth (Depth Anything V3)

How it fits

Typical backends

Caveats

• I/O contractExactly one arm is produced, fixed at deploy time by the connected consumer; fan into two instances if a graph needs both the visualisation and the metric tensor.
• Hard constraintThe repository must point at a metric (DA3METRIC) checkpoint; an any-view or mono checkpoint emits relative depth, so the metric arm aborts because no focal length is available to recover metres.
• Parameter interactionMetres are recovered from the per-frame focal length the metric head predicts; when no intrinsics are predicted it falls back to a configured pixel focal length, and otherwise to a focal derived from an assumed field of view. Scenes far from the training domain can mis-estimate the focal and rescale the whole map.
• Parameter interactionThe processing resolution caps the longest input side: smaller values speed up inference at the cost of edge detail while larger values raise VRAM and latency; the prediction is then upsampled back to native size.
• State lifetimeThe recompute cadence is configurable: unset recomputes every frame, zero computes once then frees the model and replays that map forever, and a positive value recomputes every N seconds, freeing the model between passes.
• FallbackA cuda device silently falls back to CPU with a warning when no CUDA backend is visible, continuing at CPU speeds.
• AccuracyThe visualisation Image is min-max normalised per frame, so absolute brightness is not comparable across frames; use the MetricDepthMap for any measurement or thresholding.
• State lifetimeAll configuration is captured once at startup; runtime changes have no effect and require a redeploy. Leaving the repository revision at a moving branch yields non-reproducible builds.