Estimate Metric Depth (UniDepth V2)

How it fits

Typical backends

Caveats

• I/O contractExactly one arm is produced, fixed at deploy time by the connected consumer; fan into multiple instances if a graph needs more than one of the depth, intrinsics, or point-cloud arms.
• AccuracyDepth is zero-shot absolute metric depth and the intrinsics are recovered jointly, so no calibration is needed upstream; the CameraIntrinsics arm is the full {fx, fy, cx, cy, skew, distortion} record but UniDepth recovers a pure pinhole K, so skew is zero and the distortion vector is empty; the PointCloud arm is expressed in the camera frame, not a world frame.
• Hard constraintThe UniDepth checkpoint is bound to the worker as a local artifact and loaded offline at startup; there is no download at runtime, so a cold container needs no outbound network before the first frame is served.
• State lifetimeThe recompute cadence is configurable: unset recomputes every frame, zero computes once then frees the model and replays that result forever, and a positive value recomputes every N seconds, freeing the model between passes.
• Parameter interactionThe processing-resolution level bounds the megapixel area the backbone runs at, independent of the input frame size: lower values cap GPU memory and latency so an oversized frame still fits, higher values keep more edge detail; the result is upsampled back to the native input resolution.
• CostThe ViT-L backbone strongly prefers a GPU; a cuda device silently falls back to CPU when none is visible, but CPU inference is impractically slow.
• State lifetimeAll configuration is captured once at startup; runtime changes have no effect and require a redeploy.