Run Depth Anything over tangent views

Level: Extend · Strategy: Projection-domain pretrained model · Stability: Experimental

Depth Anything 3 Metric Large is not converted to a sphere-native network in the current adapter. Its attention, patch embedding and positional treatment remain unchanged.

from panorai.experimental.deep_learning import load_da3metric_large

views = panorama.views(
    "icosahedron", subdivisions=1, size=(812, 1400), fov=(100.0, 80.0)
)
loaded = load_da3metric_large(
    accept_upstream_terms=True,
    cache_dir=model_cache,
    device="cpu",
)
adapter = loaded.tangent(views[0].spec)
predicted = views.map(adapter, input="image", output="depth", units="m")
radial = predicted.reconstruct(blend={"depth": "gaussian"}, modalities=("depth",))

PanorAi owns view geometry, virtual intrinsics, axial-to-radial conversion, support, validity, fusion and provenance. The upstream model owns the local prediction. Fusion quality at view boundaries remains part of the evaluation, not a solved property implied by model accuracy.

Why conversion is required

A pinhole model may return axial depth $z$ along its optical axis. For a face ray $d=(d_x,d_y,d_z)$ in face coordinates, the corresponding radial range is

$$\rho = z / d_z,$$

after applying the exact ray normalization and sign convention declared by the adapter. Treating $z$ as $\rho$ stretches geometry away from the face centre. The conversion is performed before spherical fusion.

What must be fused

Each face contributes radial range only where source support, model validity and finite conversion all agree. Overlap weights must not revive invalid pixels. Record the view specification and virtual intrinsics with each prediction so a seam artifact can be traced back to its source face.

Evaluation boundary

The route is an unchanged planar-model control, not a claim that the model is spherical. Evaluate view-centre versus boundary error, overlap disagreement, seam continuity, polar coverage, runtime and peak memory. Compare those results with the direct spherical CNN under the same ERP and range reference.

See the monocular-depth tutorial for the pinned artifact, terms and complete conversion contract. Next: validate the range.