1 data_handling.ase_efm.views

data_handling.ase_efm.views

Typed, zero-copy views over EFM-formatted ATEK samples.

These classes mirror the style of aria_nbv.data.views but read the keys produced by efm3d.dataset.efm_model_adaptor.load_atek_wds_dataset_as_efm. The consumed EFM key families cover ARIA_IMG/ARIA_CALIB camera streams, image/depth timestamps and frame IDs, ARIA_POSE_T_WORLD_RIG trajectories, MPS semidense point positions/uncertainties/volumes, and padded OBB/GT payloads. ATEK’s EfmModelAdaptor owns flattened-key remapping; these views only expose the resulting schema and retain the backing dictionary.

The view layer defines which fields are actor-visible (calibrated images, world-from-rig poses, and MPS semidense points) and which are oracle/evaluation assets (ARIA_OBB_PADDED, nested gt_data, and attached ASE meshes). Actor-visible EVL detected boxes enter later through backbone outputs, not EfmSnippetView.obbs. Views avoid copying large tensors so writers can attach mesh and snippet context without changing immutable VIN offline rows.

1.1 Classes

Name	Description
EfmCameraView	Expose one calibrated Aria camera stream view without copying EFM tensors.
EfmGTView	Parse nested `gt_data/efm_gt` annotations for evaluation-only access.
EfmGtCameraObbView	Expose per-camera ASE GT boxes for one EFM evaluation timestamp.
EfmGtTimestampView	Group oracle GT OBB views for one timestamp across Aria cameras.
EfmObbView	Expose snippet-level GT boxes in the EFM `ObbTW` layout.
EfmPointsView	Expose actor-visible MPS semidense reconstruction evidence.
EfmSnippetView	Own the typed boundary around one adapted ATEK/ASE snippet.
EfmTrajectoryView	Expose the MPS/EFM world-from-rig trajectory for one snippet.

1.1.1 EfmCameraView

data_handling.ase_efm.views.EfmCameraView(
    images,
    calib,
    time_ns,
    frame_ids,
    distance_m=None,
    distance_time_ns=None,
)

Expose one calibrated Aria camera stream view without copying EFM tensors.

The view consumes matching entries from ARIA_IMG, ARIA_CALIB, ARIA_IMG_TIME_NS, ARIA_FRAME_ID, and optional ARIA_DISTANCE_M/ARIA_DEPTH_TIME_NS families. VRS and online calibration are upstream provenance; runtime access is through ATEK/EFM.

1.1.1.1 Attributes

Name	Description
images	`Tensor["F C H W", float32]` normalized RGB or monochrome frames.
calib	Per-frame `CameraTW` intrinsics/extrinsics; backing shape is `(F, 34)`.
time_ns	`Tensor["F", int64]` Aria device-clock timestamps aligned to `images`.
frame_ids	`Tensor["F", int64\\|float32]` frame ids within the snippet.
distance_m	Optional `Tensor["F 1 H W", float32]` per-pixel ray distances in metres.
distance_time_ns	Optional `Tensor["F", int64]` timestamps for `distance_m`.
num_frames	Return the number of frames in the camera stream.

1.1.1.2 Methods

Name	Description
to	Move the camera view tensors to the requested device and dtype.
get_fov	Return `Tensor["F 2", float32]` horizontal/vertical FOV in degrees.
select_frame_indices	Resolve user-provided frame indices, supporting negatives and defaults.
nearest_traj_indices	Return selected camera indices and nearest trajectory indices.

1.1.1.2.1 to

data_handling.ase_efm.views.EfmCameraView.to(device, *, dtype=None)

Move the camera view tensors to the requested device and dtype.

1.1.1.2.2 get_fov

data_handling.ase_efm.views.EfmCameraView.get_fov()

Return Tensor["F 2", float32] horizontal/vertical FOV in degrees.

1.1.1.2.3 select_frame_indices

data_handling.ase_efm.views.EfmCameraView.select_frame_indices(
    frame_indices,
    *,
    default_last,
)

Resolve user-provided frame indices, supporting negatives and defaults.

1.1.1.2.4 nearest_traj_indices

data_handling.ase_efm.views.EfmCameraView.nearest_traj_indices(
    traj_ts_ns,
    frame_indices=None,
    *,
    default_last=True,
)

Return selected camera indices and nearest trajectory indices.

1.1.2 EfmGTView

data_handling.ase_efm.views.EfmGTView(raw)

Parse nested gt_data/efm_gt annotations for evaluation-only access.

1.1.2.1 Attributes

Name	Description
raw	Zero-copy backing GT payload from the adapted ATEK sample.
efm_gt	Timestamp-indexed per-camera GT views; never actor-visible observations.
timestamps	Return sorted string keys for available `efm_gt` timestamps.

1.1.2.2 Methods

Name	Description
cameras_at	Return evaluation-only per-camera GT OBBs at one timestamp.

1.1.2.2.1 cameras_at

data_handling.ase_efm.views.EfmGTView.cameras_at(ts)

Return evaluation-only per-camera GT OBBs at one timestamp.

1.1.2.2.1.1 Parameters

Name	Type	Description	Default
ts	str \| int	Timestamp key accepted as text or an integer.	required

1.1.2.2.1.2 Returns

Name	Type	Description
	dict[str, EfmGtCameraObbView]	Camera-name mapping for the requested GT slice.

1.1.3 EfmGtCameraObbView

data_handling.ase_efm.views.EfmGtCameraObbView(
    category_names,
    category_ids,
    instance_ids,
    object_dimensions,
    ts_world_object,
)

Expose per-camera ASE GT boxes for one EFM evaluation timestamp.

These tensors are label/evaluation assets used for matching and target crops. They must not be passed to an actor as observed object detections.

1.1.3.1 Attributes

Name	Description
category_names	Human-readable EFM taxonomy labels aligned with the `K` GT box rows.
category_ids	`Tensor["K", int64]` semantic IDs aligned with `category_names`.
instance_ids	`Tensor["K", int64]` GT instance IDs shared across camera views.
object_dimensions	`Tensor["K 3", float32]` object-frame box side lengths in metres.
ts_world_object	`Tensor["K 3 4", float32]` world←object GT transforms per instance.

1.1.4 EfmGtTimestampView

data_handling.ase_efm.views.EfmGtTimestampView(time_id, cameras)

Group oracle GT OBB views for one timestamp across Aria cameras.

1.1.4.1 Attributes

Name	Description
time_id	Timestamp identifier for this GT slice.
cameras	Per-camera GT OBB views keyed by camera stream name.

1.1.5 EfmObbView

data_handling.ase_efm.views.EfmObbView(obbs, hz=None)

Expose snippet-level GT boxes in the EFM ObbTW layout.

obbs may be shaped (T, K, 34), (1, K, 34), or (K, 34) depending on the source. K is padded; padded rows are all PAD_VAL and must be removed with ObbTW.get_padding_mask() before matching label targets.

The 34 payload columns are the EFM OBB contract: object-frame 3D bounds, three 2D boxes, a 12-value pose, semantic id, instance id, confidence, and movable flag. In EfmSnippetView these boxes originate from ARIA_OBB_PADDED and are GT label/evaluation data. Actor-visible target hypotheses must come from EVL predictions or tracking instead. Downstream target selection transforms the latest valid slice into world coordinates before constructing Oracle target tasks.

1.1.5.1 Attributes

Name	Description
obbs	`Tensor["... K 34", float32]` GT boxes in the source EFM frame.
hz	Optional `Tensor["1", int32]` detection frequency.

1.1.6 EfmPointsView

data_handling.ase_efm.views.EfmPointsView(
    points_world,
    dist_std,
    inv_dist_std,
    time_ns,
    volume_min,
    volume_max,
    lengths,
)

Expose actor-visible MPS semidense reconstruction evidence.

The ARIA_POINTS_* tensors are padded per frame; lengths defines the valid prefix and non-finite rows remain invalid. Positions, uncertainties, and volume bounds preserve the EFM world frame and metric units.

1.1.6.1 Attributes

Name	Description
points_world	`Tensor["F N_pad 3", float32]` padded world-frame MPS points in metres.
dist_std	`Tensor["F N_pad", float32]` per-point distance standard deviation in metres.
inv_dist_std	`Tensor["F N_pad", float32]` inverse distance uncertainty in metres⁻¹.
time_ns	`Tensor["F", int64]` timestamps aligned to point slices.
volume_min	`Tensor["3", float32]` snippet AABB minimum in world frame, metres.
volume_max	`Tensor["3", float32]` snippet AABB maximum in world frame, metres.
lengths	`Tensor["F", int64]` valid point-prefix length for each padded frame.

1.1.6.2 Methods

Name	Description
to	Move the semidense point tensors to the requested device and dtype.
collapse_points	Collapse points across time and optionally subsample.
last_frame_points_np	Return points from the latest frame with valid points, subsampled if needed.

1.1.6.2.1 to

data_handling.ase_efm.views.EfmPointsView.to(device, *, dtype=None)

Move the semidense point tensors to the requested device and dtype.

1.1.6.2.2 collapse_points

data_handling.ase_efm.views.EfmPointsView.collapse_points(
    max_points=None,
    include_inv_dist_std=False,
    include_obs_count=False,
)

Collapse points across time and optionally subsample.

1.1.6.2.2.1 Parameters

Name	Type	Description	Default
max_points	int \| None	Optional cap on the number of returned points.	`None`
include_inv_dist_std	bool	If True, append inverse depth std per point.	`False`
include_obs_count	bool	If True, append the number of snippet frames that observed each point (observation count).	`False`

1.1.6.2.2.2 Returns

Name	Type	Description
	Tensor	`Tensor["K 3", float32]` if no extra features are requested.
	Tensor	`Tensor["K 4", float32]` if exactly one extra feature is requested
	Tensor	(XYZ + inv_dist_std or XYZ + obs_count).
	Tensor	`Tensor["K 5", float32]` if both extras are requested
	Tensor	(XYZ + inv_dist_std + obs_count).

1.1.6.2.3 last_frame_points_np

data_handling.ase_efm.views.EfmPointsView.last_frame_points_np(max_points=None)

Return points from the latest frame with valid points, subsampled if needed.

1.1.7 EfmSnippetView

data_handling.ase_efm.views.EfmSnippetView(
    efm,
    scene_id,
    snippet_id,
    mesh=None,
    crop_bounds=None,
    mesh_verts=None,
    mesh_faces=None,
    mesh_cache_key=None,
    mesh_specs=None,
)

Own the typed boundary around one adapted ATEK/ASE snippet.

efm retains the zero-copy key/value payload produced by EfmModelAdaptor. Camera/calibration, world-from-rig trajectory, and MPS semidense fields are actor-visible observations. Attached ASE meshes, ARIA_OBB_PADDED, and nested gt_data are oracle/evaluation assets. Mesh tensors are optional CPU/cache products and are not moved into the EFM dictionary.

1.1.7.1 Attributes

Name	Description
efm	Zero-copy adapted ATEK payload containing the consumed `ARIA_*` key families.
scene_id	ASE scene identifier used for shard lineage and GT-mesh pairing.
snippet_id	Compact ATEK sample identifier derived from the WebDataset sample key.
mesh	Optional ASE GT mesh used only by oracle labeling and evaluation.
crop_bounds	Optional pair of `Tensor["3", float32]` world-frame bounds in metres.
mesh_verts	Optional `Tensor["V 3", float32]` GT-mesh vertices in world frame, metres.
mesh_faces	Optional `Tensor["F_mesh 3", int64]` triangle vertex indices.
mesh_cache_key	Stable key (spec hash) for shared mesh caches across components.
mesh_specs	Optional mesh processing specification used to derive attached mesh tensors.
camera_rgb	RGB stream (fisheye624, 240x240 in ASE preprocessing).
camera_slam_left	Left SLAM mono stream (fisheye624, ~320x240).
camera_slam_right	Right SLAM mono stream (fisheye624, ~320x240).
trajectory	Rig trajectory (world←rig) aligned to snippet frames.
semidense	Semi-dense SLAM points (padded to fixed length).
obbs	Return padded ASE GT OBBs for label/evaluation use when present.
gt	Ground-truth dict `gt_data` (EFM OBBs per timestamp/camera).
has_mesh	Return whether a ground-truth mesh is attached to the snippet.

1.1.7.2 Methods

Name	Description
from_cache_efm	Construct a snippet view from an offline-cache EFM dict.
get_camera	Return the requested camera stream view from the backing EFM dict.
get_occupancy_extend	Return `[xmin, xmax, ymin, ymax, zmin, zmax]` AABB in world frame.
to	Return a shallow copy with EFM tensors and wrappers moved to a device.
prune_efm	Return a view with the EFM dict pruned to the requested keys.

1.1.7.2.1 from_cache_efm

data_handling.ase_efm.views.EfmSnippetView.from_cache_efm(
    efm,
    *,
    mesh=None,
    crop_bounds=None,
    mesh_verts=None,
    mesh_faces=None,
    mesh_cache_key=None,
    mesh_specs=None,
)

Construct a snippet view from an offline-cache EFM dict.

1.1.7.2.1.1 Parameters

Name	Type	Description	Default
efm	dict[str, Any]	Raw cache sample dict that includes `"__key__"`.	required
mesh	`Trimesh` \| None	Optional GT mesh to attach.	`None`
crop_bounds	tuple[torch.Tensor, torch.Tensor] \| None	Optional world-space AABB override.	`None`
mesh_verts	torch.Tensor \| None	Optional cached mesh vertices.	`None`
mesh_faces	torch.Tensor \| None	Optional cached mesh faces.	`None`
mesh_cache_key	str \| None	Optional mesh cache key for shared caches.	`None`
mesh_specs	MeshProcessSpec \| None	Optional mesh processing spec.	`None`

1.1.7.2.1.2 Returns

Name	Type	Description
	'EfmSnippetView'	Parsed `EfmSnippetView` instance.

1.1.7.2.2 get_camera

data_handling.ase_efm.views.EfmSnippetView.get_camera(prefix)

Return the requested camera stream view from the backing EFM dict.

1.1.7.2.3 get_occupancy_extend

data_handling.ase_efm.views.EfmSnippetView.get_occupancy_extend()

Return [xmin, xmax, ymin, ymax, zmin, zmax] AABB in world frame.

1.1.7.2.4 to

data_handling.ase_efm.views.EfmSnippetView.to(device, *, dtype=None)

Return a shallow copy with EFM tensors and wrappers moved to a device.

The backing dictionary is rebuilt without mutating this view. Plain tensors honor dtype; EFM wrappers preserve their own dtype contract. The CPU Trimesh object is shared, while cached mesh tensors and world-frame crop bounds are moved.

1.1.7.2.5 prune_efm

data_handling.ase_efm.views.EfmSnippetView.prune_efm(
    keep_keys,
    *,
    keep_prefixes=(),
)

Return a view with the EFM dict pruned to the requested keys.

1.1.7.2.5.1 Parameters

Name	Type	Description	Default
keep_keys	set[str] \| None	Exact EFM keys to retain. `None` returns `self` unchanged.	required
keep_prefixes	tuple[str, …]	Optional prefixes; keys starting with any prefix are retained.	`()`

1.1.8 EfmTrajectoryView

data_handling.ase_efm.views.EfmTrajectoryView(
    t_world_rig,
    time_ns,
    gravity_in_world,
)

Expose the MPS/EFM world-from-rig trajectory for one snippet.

1.1.8.1 Attributes

Name	Description
t_world_rig	`PoseTW["F 12"]` world←rig transforms; translation is metres.
time_ns	`Tensor["F", int64]` pose timestamps.
gravity_in_world	`Tensor["3", float32]` world-frame gravity in metres per second squared.
final_pose	Return the final `PoseTW["12"]` world←rig pose in the snippet.

1.1.8.2 Methods

Name	Description
to	Move the trajectory tensors to the requested device and dtype.

1.1.8.2.1 to

data_handling.ase_efm.views.EfmTrajectoryView.to(device, *, dtype=None)

Move the trajectory tensors to the requested device and dtype.

1.2 Functions

Name	Description
is_efm_snippet_view_instance	Return whether `value` behaves like an `EfmSnippetView`.

1.2.1 is_efm_snippet_view_instance

data_handling.ase_efm.views.is_efm_snippet_view_instance(value)

Return whether value behaves like an EfmSnippetView.

The v2 stack accepts both the local data-handling view classes and legacy view objects that expose the same public attributes.

# data_handling.ase_efm.views { #aria_nbv.data_handling.ase_efm.views } `data_handling.ase_efm.views` Typed, zero-copy views over EFM-formatted ATEK samples. These classes mirror the style of `aria_nbv.data.views` but read the keys produced by ``efm3d.dataset.efm_model_adaptor.load_atek_wds_dataset_as_efm``. The consumed EFM key families cover ``ARIA_IMG``/``ARIA_CALIB`` camera streams, image/depth timestamps and frame IDs, ``ARIA_POSE_T_WORLD_RIG`` trajectories, MPS semidense point positions/uncertainties/volumes, and padded OBB/GT payloads. ATEK's ``EfmModelAdaptor`` owns flattened-key remapping; these views only expose the resulting schema and retain the backing dictionary. The view layer defines which fields are actor-visible (calibrated images, world-from-rig poses, and MPS semidense points) and which are oracle/evaluation assets (``ARIA_OBB_PADDED``, nested ``gt_data``, and attached ASE meshes). Actor-visible EVL detected boxes enter later through backbone outputs, not :attr:`EfmSnippetView.obbs`. Views avoid copying large tensors so writers can attach mesh and snippet context without changing immutable VIN offline rows. ## Classes | Name | Description | | --- | --- | | [EfmCameraView](#aria_nbv.data_handling.ase_efm.views.EfmCameraView) | Expose one calibrated Aria camera stream view without copying EFM tensors. | | [EfmGTView](#aria_nbv.data_handling.ase_efm.views.EfmGTView) | Parse nested ``gt_data/efm_gt`` annotations for evaluation-only access. | | [EfmGtCameraObbView](#aria_nbv.data_handling.ase_efm.views.EfmGtCameraObbView) | Expose per-camera ASE GT boxes for one EFM evaluation timestamp. | | [EfmGtTimestampView](#aria_nbv.data_handling.ase_efm.views.EfmGtTimestampView) | Group oracle GT OBB views for one timestamp across Aria cameras. | | [EfmObbView](#aria_nbv.data_handling.ase_efm.views.EfmObbView) | Expose snippet-level GT boxes in the EFM ``ObbTW`` layout. | | [EfmPointsView](#aria_nbv.data_handling.ase_efm.views.EfmPointsView) | Expose actor-visible MPS semidense reconstruction evidence. | | [EfmSnippetView](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView) | Own the typed boundary around one adapted ATEK/ASE snippet. | | [EfmTrajectoryView](#aria_nbv.data_handling.ase_efm.views.EfmTrajectoryView) | Expose the MPS/EFM world-from-rig trajectory for one snippet. | ### EfmCameraView { #aria_nbv.data_handling.ase_efm.views.EfmCameraView } ```python data_handling.ase_efm.views.EfmCameraView( images, calib, time_ns, frame_ids, distance_m=None, distance_time_ns=None, ) ``` Expose one calibrated Aria camera stream view without copying EFM tensors. The view consumes matching entries from ``ARIA_IMG``, ``ARIA_CALIB``, ``ARIA_IMG_TIME_NS``, ``ARIA_FRAME_ID``, and optional ``ARIA_DISTANCE_M``/``ARIA_DEPTH_TIME_NS`` families. VRS and online calibration are upstream provenance; runtime access is through ATEK/EFM. #### Attributes | Name | Description | | --- | --- | | [images](#aria_nbv.data_handling.ase_efm.views.EfmCameraView.images) | ``Tensor["F C H W", float32]`` normalized RGB or monochrome frames. | | [calib](#aria_nbv.data_handling.ase_efm.views.EfmCameraView.calib) | Per-frame ``CameraTW`` intrinsics/extrinsics; backing shape is ``(F, 34)``. | | [time_ns](#aria_nbv.data_handling.ase_efm.views.EfmCameraView.time_ns) | ``Tensor["F", int64]`` Aria device-clock timestamps aligned to ``images``. | | [frame_ids](#aria_nbv.data_handling.ase_efm.views.EfmCameraView.frame_ids) | ``Tensor["F", int64\|float32]`` frame ids within the snippet. | | [distance_m](#aria_nbv.data_handling.ase_efm.views.EfmCameraView.distance_m) | Optional ``Tensor["F 1 H W", float32]`` per-pixel ray distances in metres. | | [distance_time_ns](#aria_nbv.data_handling.ase_efm.views.EfmCameraView.distance_time_ns) | Optional ``Tensor["F", int64]`` timestamps for ``distance_m``. | | [num_frames](#aria_nbv.data_handling.ase_efm.views.EfmCameraView.num_frames) | Return the number of frames in the camera stream. | #### Methods | Name | Description | | --- | --- | | [to](#aria_nbv.data_handling.ase_efm.views.EfmCameraView.to) | Move the camera view tensors to the requested device and dtype. | | [get_fov](#aria_nbv.data_handling.ase_efm.views.EfmCameraView.get_fov) | Return ``Tensor["F 2", float32]`` horizontal/vertical FOV in degrees. | | [select_frame_indices](#aria_nbv.data_handling.ase_efm.views.EfmCameraView.select_frame_indices) | Resolve user-provided frame indices, supporting negatives and defaults. | | [nearest_traj_indices](#aria_nbv.data_handling.ase_efm.views.EfmCameraView.nearest_traj_indices) | Return selected camera indices and nearest trajectory indices. | ##### to { #aria_nbv.data_handling.ase_efm.views.EfmCameraView.to } ```python data_handling.ase_efm.views.EfmCameraView.to(device, *, dtype=None) ``` Move the camera view tensors to the requested device and dtype. ##### get_fov { #aria_nbv.data_handling.ase_efm.views.EfmCameraView.get_fov } ```python data_handling.ase_efm.views.EfmCameraView.get_fov() ``` Return ``Tensor["F 2", float32]`` horizontal/vertical FOV in degrees. ##### select_frame_indices { #aria_nbv.data_handling.ase_efm.views.EfmCameraView.select_frame_indices } ```python data_handling.ase_efm.views.EfmCameraView.select_frame_indices( frame_indices, *, default_last, ) ``` Resolve user-provided frame indices, supporting negatives and defaults. ##### nearest_traj_indices { #aria_nbv.data_handling.ase_efm.views.EfmCameraView.nearest_traj_indices } ```python data_handling.ase_efm.views.EfmCameraView.nearest_traj_indices( traj_ts_ns, frame_indices=None, *, default_last=True, ) ``` Return selected camera indices and nearest trajectory indices. ### EfmGTView { #aria_nbv.data_handling.ase_efm.views.EfmGTView } ```python data_handling.ase_efm.views.EfmGTView(raw) ``` Parse nested ``gt_data/efm_gt`` annotations for evaluation-only access. #### Attributes | Name | Description | | --- | --- | | [raw](#aria_nbv.data_handling.ase_efm.views.EfmGTView.raw) | Zero-copy backing GT payload from the adapted ATEK sample. | | [efm_gt](#aria_nbv.data_handling.ase_efm.views.EfmGTView.efm_gt) | Timestamp-indexed per-camera GT views; never actor-visible observations. | | [timestamps](#aria_nbv.data_handling.ase_efm.views.EfmGTView.timestamps) | Return sorted string keys for available ``efm_gt`` timestamps. | #### Methods | Name | Description | | --- | --- | | [cameras_at](#aria_nbv.data_handling.ase_efm.views.EfmGTView.cameras_at) | Return evaluation-only per-camera GT OBBs at one timestamp. | ##### cameras_at { #aria_nbv.data_handling.ase_efm.views.EfmGTView.cameras_at } ```python data_handling.ase_efm.views.EfmGTView.cameras_at(ts) ``` Return evaluation-only per-camera GT OBBs at one timestamp. ###### Parameters {.doc-section .doc-section-parameters} | Name | Type | Description | Default | |--------|------------------------------|-----------------------------------------------|------------| | ts | [str](`str`) \| [int](`int`) | Timestamp key accepted as text or an integer. | _required_ | ###### Returns {.doc-section .doc-section-returns} | Name | Type | Description | |--------|-----------------------------------------------------------------------------------------------------------------|-------------------------------------------------| | | [dict](`dict`)\[[str](`str`), [EfmGtCameraObbView](`aria_nbv.data_handling.ase_efm.views.EfmGtCameraObbView`)\] | Camera-name mapping for the requested GT slice. | ### EfmGtCameraObbView { #aria_nbv.data_handling.ase_efm.views.EfmGtCameraObbView } ```python data_handling.ase_efm.views.EfmGtCameraObbView( category_names, category_ids, instance_ids, object_dimensions, ts_world_object, ) ``` Expose per-camera ASE GT boxes for one EFM evaluation timestamp. These tensors are label/evaluation assets used for matching and target crops. They must not be passed to an actor as observed object detections. #### Attributes | Name | Description | | --- | --- | | [category_names](#aria_nbv.data_handling.ase_efm.views.EfmGtCameraObbView.category_names) | Human-readable EFM taxonomy labels aligned with the ``K`` GT box rows. | | [category_ids](#aria_nbv.data_handling.ase_efm.views.EfmGtCameraObbView.category_ids) | ``Tensor["K", int64]`` semantic IDs aligned with ``category_names``. | | [instance_ids](#aria_nbv.data_handling.ase_efm.views.EfmGtCameraObbView.instance_ids) | ``Tensor["K", int64]`` GT instance IDs shared across camera views. | | [object_dimensions](#aria_nbv.data_handling.ase_efm.views.EfmGtCameraObbView.object_dimensions) | ``Tensor["K 3", float32]`` object-frame box side lengths in metres. | | [ts_world_object](#aria_nbv.data_handling.ase_efm.views.EfmGtCameraObbView.ts_world_object) | ``Tensor["K 3 4", float32]`` world←object GT transforms per instance. | ### EfmGtTimestampView { #aria_nbv.data_handling.ase_efm.views.EfmGtTimestampView } ```python data_handling.ase_efm.views.EfmGtTimestampView(time_id, cameras) ``` Group oracle GT OBB views for one timestamp across Aria cameras. #### Attributes | Name | Description | | --- | --- | | [time_id](#aria_nbv.data_handling.ase_efm.views.EfmGtTimestampView.time_id) | Timestamp identifier for this GT slice. | | [cameras](#aria_nbv.data_handling.ase_efm.views.EfmGtTimestampView.cameras) | Per-camera GT OBB views keyed by camera stream name. | ### EfmObbView { #aria_nbv.data_handling.ase_efm.views.EfmObbView } ```python data_handling.ase_efm.views.EfmObbView(obbs, hz=None) ``` Expose snippet-level GT boxes in the EFM ``ObbTW`` layout. `obbs` may be shaped `(T, K, 34)`, `(1, K, 34)`, or `(K, 34)` depending on the source. ``K`` is padded; padded rows are all ``PAD_VAL`` and must be removed with ``ObbTW.get_padding_mask()`` before matching label targets. The `34` payload columns are the EFM OBB contract: object-frame 3D bounds, three 2D boxes, a 12-value pose, semantic id, instance id, confidence, and movable flag. In :class:`EfmSnippetView` these boxes originate from ``ARIA_OBB_PADDED`` and are GT label/evaluation data. Actor-visible target hypotheses must come from EVL predictions or tracking instead. Downstream target selection transforms the latest valid slice into world coordinates before constructing Oracle target tasks. #### Attributes | Name | Description | | --- | --- | | [obbs](#aria_nbv.data_handling.ase_efm.views.EfmObbView.obbs) | ``Tensor["... K 34", float32]`` GT boxes in the source EFM frame. | | [hz](#aria_nbv.data_handling.ase_efm.views.EfmObbView.hz) | Optional ``Tensor["1", int32]`` detection frequency. | ### EfmPointsView { #aria_nbv.data_handling.ase_efm.views.EfmPointsView } ```python data_handling.ase_efm.views.EfmPointsView( points_world, dist_std, inv_dist_std, time_ns, volume_min, volume_max, lengths, ) ``` Expose actor-visible MPS semidense reconstruction evidence. The ``ARIA_POINTS_*`` tensors are padded per frame; ``lengths`` defines the valid prefix and non-finite rows remain invalid. Positions, uncertainties, and volume bounds preserve the EFM world frame and metric units. #### Attributes | Name | Description | | --- | --- | | [points_world](#aria_nbv.data_handling.ase_efm.views.EfmPointsView.points_world) | ``Tensor["F N_pad 3", float32]`` padded world-frame MPS points in metres. | | [dist_std](#aria_nbv.data_handling.ase_efm.views.EfmPointsView.dist_std) | ``Tensor["F N_pad", float32]`` per-point distance standard deviation in metres. | | [inv_dist_std](#aria_nbv.data_handling.ase_efm.views.EfmPointsView.inv_dist_std) | ``Tensor["F N_pad", float32]`` inverse distance uncertainty in metres⁻¹. | | [time_ns](#aria_nbv.data_handling.ase_efm.views.EfmPointsView.time_ns) | ``Tensor["F", int64]`` timestamps aligned to point slices. | | [volume_min](#aria_nbv.data_handling.ase_efm.views.EfmPointsView.volume_min) | ``Tensor["3", float32]`` snippet AABB minimum in world frame, metres. | | [volume_max](#aria_nbv.data_handling.ase_efm.views.EfmPointsView.volume_max) | ``Tensor["3", float32]`` snippet AABB maximum in world frame, metres. | | [lengths](#aria_nbv.data_handling.ase_efm.views.EfmPointsView.lengths) | ``Tensor["F", int64]`` valid point-prefix length for each padded frame. | #### Methods | Name | Description | | --- | --- | | [to](#aria_nbv.data_handling.ase_efm.views.EfmPointsView.to) | Move the semidense point tensors to the requested device and dtype. | | [collapse_points](#aria_nbv.data_handling.ase_efm.views.EfmPointsView.collapse_points) | Collapse points across time and optionally subsample. | | [last_frame_points_np](#aria_nbv.data_handling.ase_efm.views.EfmPointsView.last_frame_points_np) | Return points from the latest frame with valid points, subsampled if needed. | ##### to { #aria_nbv.data_handling.ase_efm.views.EfmPointsView.to } ```python data_handling.ase_efm.views.EfmPointsView.to(device, *, dtype=None) ``` Move the semidense point tensors to the requested device and dtype. ##### collapse_points { #aria_nbv.data_handling.ase_efm.views.EfmPointsView.collapse_points } ```python data_handling.ase_efm.views.EfmPointsView.collapse_points( max_points=None, include_inv_dist_std=False, include_obs_count=False, ) ``` Collapse points across time and optionally subsample. ###### Parameters {.doc-section .doc-section-parameters} | Name | Type | Description | Default | |----------------------|----------------------|--------------------------------------------------------------------------------------------|-----------| | max_points | [int](`int`) \| None | Optional cap on the number of returned points. | `None` | | include_inv_dist_std | [bool](`bool`) | If True, append inverse depth std per point. | `False` | | include_obs_count | [bool](`bool`) | If True, append the number of snippet frames that observed each point (observation count). | `False` | ###### Returns {.doc-section .doc-section-returns} | Name | Type | Description | |--------|--------------------------|----------------------------------------------------------------------| | | [Tensor](`torch.Tensor`) | ``Tensor["K 3", float32]`` if no extra features are requested. | | | [Tensor](`torch.Tensor`) | ``Tensor["K 4", float32]`` if exactly one extra feature is requested | | | [Tensor](`torch.Tensor`) | (XYZ + inv_dist_std or XYZ + obs_count). | | | [Tensor](`torch.Tensor`) | ``Tensor["K 5", float32]`` if both extras are requested | | | [Tensor](`torch.Tensor`) | (XYZ + inv_dist_std + obs_count). | ##### last_frame_points_np { #aria_nbv.data_handling.ase_efm.views.EfmPointsView.last_frame_points_np } ```python data_handling.ase_efm.views.EfmPointsView.last_frame_points_np(max_points=None) ``` Return points from the latest frame with valid points, subsampled if needed. ### EfmSnippetView { #aria_nbv.data_handling.ase_efm.views.EfmSnippetView } ```python data_handling.ase_efm.views.EfmSnippetView( efm, scene_id, snippet_id, mesh=None, crop_bounds=None, mesh_verts=None, mesh_faces=None, mesh_cache_key=None, mesh_specs=None, ) ``` Own the typed boundary around one adapted ATEK/ASE snippet. ``efm`` retains the zero-copy key/value payload produced by ``EfmModelAdaptor``. Camera/calibration, world-from-rig trajectory, and MPS semidense fields are actor-visible observations. Attached ASE meshes, ``ARIA_OBB_PADDED``, and nested ``gt_data`` are oracle/evaluation assets. Mesh tensors are optional CPU/cache products and are not moved into the EFM dictionary. #### Attributes | Name | Description | | --- | --- | | [efm](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.efm) | Zero-copy adapted ATEK payload containing the consumed ``ARIA_*`` key families. | | [scene_id](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.scene_id) | ASE scene identifier used for shard lineage and GT-mesh pairing. | | [snippet_id](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.snippet_id) | Compact ATEK sample identifier derived from the WebDataset sample key. | | [mesh](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.mesh) | Optional ASE GT mesh used only by oracle labeling and evaluation. | | [crop_bounds](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.crop_bounds) | Optional pair of ``Tensor["3", float32]`` world-frame bounds in metres. | | [mesh_verts](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.mesh_verts) | Optional ``Tensor["V 3", float32]`` GT-mesh vertices in world frame, metres. | | [mesh_faces](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.mesh_faces) | Optional ``Tensor["F_mesh 3", int64]`` triangle vertex indices. | | [mesh_cache_key](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.mesh_cache_key) | Stable key (spec hash) for shared mesh caches across components. | | [mesh_specs](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.mesh_specs) | Optional mesh processing specification used to derive attached mesh tensors. | | [camera_rgb](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.camera_rgb) | RGB stream (fisheye624, 240x240 in ASE preprocessing). | | [camera_slam_left](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.camera_slam_left) | Left SLAM mono stream (fisheye624, ~320x240). | | [camera_slam_right](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.camera_slam_right) | Right SLAM mono stream (fisheye624, ~320x240). | | [trajectory](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.trajectory) | Rig trajectory (world←rig) aligned to snippet frames. | | [semidense](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.semidense) | Semi-dense SLAM points (padded to fixed length). | | [obbs](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.obbs) | Return padded ASE GT OBBs for label/evaluation use when present. | | [gt](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.gt) | Ground-truth dict ``gt_data`` (EFM OBBs per timestamp/camera). | | [has_mesh](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.has_mesh) | Return whether a ground-truth mesh is attached to the snippet. | #### Methods | Name | Description | | --- | --- | | [from_cache_efm](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.from_cache_efm) | Construct a snippet view from an offline-cache EFM dict. | | [get_camera](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.get_camera) | Return the requested camera stream view from the backing EFM dict. | | [get_occupancy_extend](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.get_occupancy_extend) | Return ``[xmin, xmax, ymin, ymax, zmin, zmax]`` AABB in world frame. | | [to](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.to) | Return a shallow copy with EFM tensors and wrappers moved to a device. | | [prune_efm](#aria_nbv.data_handling.ase_efm.views.EfmSnippetView.prune_efm) | Return a view with the EFM dict pruned to the requested keys. | ##### from_cache_efm { #aria_nbv.data_handling.ase_efm.views.EfmSnippetView.from_cache_efm } ```python data_handling.ase_efm.views.EfmSnippetView.from_cache_efm( efm, *, mesh=None, crop_bounds=None, mesh_verts=None, mesh_faces=None, mesh_cache_key=None, mesh_specs=None, ) ``` Construct a snippet view from an offline-cache EFM dict. ###### Parameters {.doc-section .doc-section-parameters} | Name | Type | Description | Default | |----------------|------------------------------------------------------------------------------------------------------------------|----------------------------------------------------|------------| | efm | [dict](`dict`)\[[str](`str`), [Any](`typing.Any`)\] | Raw cache sample dict that includes ``"__key__"``. | _required_ | | mesh | [Trimesh](`trimesh.Trimesh`) \| None | Optional GT mesh to attach. | `None` | | crop_bounds | [tuple](`tuple`)\[[torch](`torch`).[Tensor](`torch.Tensor`), [torch](`torch`).[Tensor](`torch.Tensor`)\] \| None | Optional world-space AABB override. | `None` | | mesh_verts | [torch](`torch`).[Tensor](`torch.Tensor`) \| None | Optional cached mesh vertices. | `None` | | mesh_faces | [torch](`torch`).[Tensor](`torch.Tensor`) \| None | Optional cached mesh faces. | `None` | | mesh_cache_key | [str](`str`) \| None | Optional mesh cache key for shared caches. | `None` | | mesh_specs | [MeshProcessSpec](`aria_nbv.data_handling.mesh_cache.MeshProcessSpec`) \| None | Optional mesh processing spec. | `None` | ###### Returns {.doc-section .doc-section-returns} | Name | Type | Description | |--------|--------------------|-----------------------------------| | | \'EfmSnippetView\' | Parsed `EfmSnippetView` instance. | ##### get_camera { #aria_nbv.data_handling.ase_efm.views.EfmSnippetView.get_camera } ```python data_handling.ase_efm.views.EfmSnippetView.get_camera(prefix) ``` Return the requested camera stream view from the backing EFM dict. ##### get_occupancy_extend { #aria_nbv.data_handling.ase_efm.views.EfmSnippetView.get_occupancy_extend } ```python data_handling.ase_efm.views.EfmSnippetView.get_occupancy_extend() ``` Return ``[xmin, xmax, ymin, ymax, zmin, zmax]`` AABB in world frame. ##### to { #aria_nbv.data_handling.ase_efm.views.EfmSnippetView.to } ```python data_handling.ase_efm.views.EfmSnippetView.to(device, *, dtype=None) ``` Return a shallow copy with EFM tensors and wrappers moved to a device. The backing dictionary is rebuilt without mutating this view. Plain tensors honor ``dtype``; EFM wrappers preserve their own dtype contract. The CPU ``Trimesh`` object is shared, while cached mesh tensors and world-frame crop bounds are moved. ##### prune_efm { #aria_nbv.data_handling.ase_efm.views.EfmSnippetView.prune_efm } ```python data_handling.ase_efm.views.EfmSnippetView.prune_efm( keep_keys, *, keep_prefixes=(), ) ``` Return a view with the EFM dict pruned to the requested keys. ###### Parameters {.doc-section .doc-section-parameters} | Name | Type | Description | Default | |---------------|---------------------------------------|----------------------------------------------------------------|------------| | keep_keys | [set](`set`)\[[str](`str`)\] \| None | Exact EFM keys to retain. ``None`` returns ``self`` unchanged. | _required_ | | keep_prefixes | [tuple](`tuple`)\[[str](`str`), ...\] | Optional prefixes; keys starting with any prefix are retained. | `()` | ### EfmTrajectoryView { #aria_nbv.data_handling.ase_efm.views.EfmTrajectoryView } ```python data_handling.ase_efm.views.EfmTrajectoryView( t_world_rig, time_ns, gravity_in_world, ) ``` Expose the MPS/EFM world-from-rig trajectory for one snippet. #### Attributes | Name | Description | | --- | --- | | [t_world_rig](#aria_nbv.data_handling.ase_efm.views.EfmTrajectoryView.t_world_rig) | ``PoseTW["F 12"]`` world←rig transforms; translation is metres. | | [time_ns](#aria_nbv.data_handling.ase_efm.views.EfmTrajectoryView.time_ns) | ``Tensor["F", int64]`` pose timestamps. | | [gravity_in_world](#aria_nbv.data_handling.ase_efm.views.EfmTrajectoryView.gravity_in_world) | ``Tensor["3", float32]`` world-frame gravity in metres per second squared. | | [final_pose](#aria_nbv.data_handling.ase_efm.views.EfmTrajectoryView.final_pose) | Return the final ``PoseTW["12"]`` world←rig pose in the snippet. | #### Methods | Name | Description | | --- | --- | | [to](#aria_nbv.data_handling.ase_efm.views.EfmTrajectoryView.to) | Move the trajectory tensors to the requested device and dtype. | ##### to { #aria_nbv.data_handling.ase_efm.views.EfmTrajectoryView.to } ```python data_handling.ase_efm.views.EfmTrajectoryView.to(device, *, dtype=None) ``` Move the trajectory tensors to the requested device and dtype. ## Functions | Name | Description | | --- | --- | | [is_efm_snippet_view_instance](#aria_nbv.data_handling.ase_efm.views.is_efm_snippet_view_instance) | Return whether ``value`` behaves like an `EfmSnippetView`. | ### is_efm_snippet_view_instance { #aria_nbv.data_handling.ase_efm.views.is_efm_snippet_view_instance } ```python data_handling.ase_efm.views.is_efm_snippet_view_instance(value) ``` Return whether ``value`` behaves like an `EfmSnippetView`. The v2 stack accepts both the local data-handling view classes and legacy view objects that expose the same public attributes.