Estimate Metric Depth (Depth Anything V2)

@bytedance1 VersionEstimate

Metric monocular depth from Depth Anything V2: turns each Image into a deploy-time-resolved greyscale depth Image or a MetricDepthMap in metres (near small, far large), capped to a fine-tuned maximum range.

How it fits

Typical backends

A live image stream is turned into a normalised greyscale depth view and streamed to the browser for inspection.

Metric depth from recorded frames is shipped as structured data over HTTP for downstream measurement or logging.

Depth is expensive and changes slowly, so a gate thins the live stream to a sparse cadence the GPU recomputes on, and a fuse holds that slow depth map and re-presents it alongside every live frame at the full input rate.

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 encoder backbone name accepts only vits, vitb, vitl, or vitg; any other value aborts startup, and the encoder size must match the architecture of the checkpoint weights or the network loads silently and produces garbage depth.
Hard constraintThe checkpoint is bound to the worker as a local .pth/.pt artifact and loaded offline at startup; there is no download at runtime, and exactly one checkpoint file must be bound.
Parameter interactionThe metric cap in metres must match the cap the checkpoint was fine-tuned with; a mismatched value silently re-scales the whole depth map.
State lifetimeWith release_memory true (the default) the model and CUDA cache are freed after each inference and lazily reloaded on the next frame; the worker runs depth on every frame it receives, so thin the input with an upstream gate and hold the slow map at the upstream rate with a downstream fuse. Set it false to keep the model resident for lowest latency when depth runs every frame.
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.

Versions

064b1829latestdefault
Automated release
6/30/2026