1 vin.types.volume

vin.types.volume

Typed deterministic evidence carried by the volumetric VIN representation.

1.1 Attributes

Name Description
ROOT_VOLUME_CHANNELS Stable ordering of immutable root channels in VolumeEvidence.root_features.
VOLUME_SCENE_CHANNELS Stable network-channel ordering returned by VolumeEvidence.scene_tensor.

1.2 Classes

Name Description
VolumeEvidence Fixed root-aligned evidence grid for one batch of rollout chains.

1.2.1 VolumeEvidence

vin.types.volume.VolumeEvidence(
    root_features,
    root_presence,
    evl_support,
    root_observed_support,
    selected_surface_count,
    selected_free_count,
    last_observed_step,
    t_world_volume,
    volume_extent,
    integrated_observation_ids,
    count_normalization,
    recency_normalization,
    depth_noise_m,
)

Fixed root-aligned evidence grid for one batch of rollout chains.

Immutable root priors remain separate from selected-depth measurements. In particular, an EVL occupancy prediction is not measurement support, and the expanded part of the map is not treated as observed merely because the dense tensor has materialized cells there.

The grid layout is B,C,D,H,W. Metric x, y, and z axes map to W, H, and D respectively. t_world_volume stores the fixed world-from-volume chart pose and volume_extent stores its metric bounds in volume coordinates.

1.2.1.1 Attributes

Name Description
root_features Tensor["B 6 D H W", float32] root channels ordered by ROOT_VOLUME_CHANNELS.
root_presence Tensor["B 6", bool] source availability; numeric zero never substitutes for absence.
evl_support Tensor["B 1 D H W", bool] cells whose centres lie inside the original EVL chart.
root_observed_support Tensor["B 1 D H W", bool] measured root occupied/count or semidense support.
selected_surface_count Tensor["B 1 D H W", float32] capped one-vote-per-observation surface counts.
selected_free_count Tensor["B 1 D H W", float32] capped one-vote-per-observation free-space counts.
last_observed_step Tensor["B 1 D H W", float32] latest selected observation id, or -1 if unseen.
t_world_volume PoseTW["B 12"] world-from-volume pose, fixed for the chain.
volume_extent Tensor["B 6", float32] expanded bounds [xmin,xmax,ymin,ymax,zmin,zmax] in metres.
integrated_observation_ids Per-chain selected observation ids already fused into the dynamic channels.
count_normalization Fixed positive count scale used by log1p network channels.
recency_normalization Fixed positive observation-index scale used by the recency channel.
depth_noise_m Non-negative camera-z noise allowance added to the voxel-size surface band.
observed_support Return explicit measured support, distinct from prediction and map materialization.
scene_tensor Return the stable seventeen-channel float32 tensor consumed by the volume encoder.
Back to top