ATEK Implementation Index

1 Purpose

ATEK is the production data pipeline for Aria Synthetic Environments. This index lists signatures, shapes, and theory for building/flattening/loading snippets and for evaluation utilities we depend on (plus items we underuse).

1.1 Wikipedia theory primer

SLAM (Simultaneous Localization and Mapping): joint estimation of a sensor’s pose and a map of the environment from the same observations—classically solved with probabilistic filters combining odometry and landmark updates. ATEK’s trajectory + semidense streams are the outputs of this SLAM stage.
Source: Wikipedia — Simultaneous localization and mapping.
SE(3) pose representation: Special Euclidean group combining 3D rotations (SO(3)) and translations; pose composition is associative and invertible, which keeps T_world_device @ T_device_camera well-defined for every frame.
Sources: Special Euclidean group and Special orthogonal group.

2 File-by-file implementation guide (ATEK)

data_preprocess/atek_data_sample.py
Nested dataclasses for camera, trajectory, semidense, online calib, and GT containers; to_flatten_dict/create_atek_data_sample_from_flatten_dict handle stable WDS key naming and reconstruction.
processors/aria_camera_processor.py
Reads VRS streams, extracts per-frame calibration, applies pixel transforms, and outputs MultiFrameCameraData; supports label-dependent resize/crop.
processors/depth_image_processor.py
Loads depth VRS, converts z-depth to distance, precomputes pixel-to-ray cache, and aligns timestamps to RGB/trajectory; returns [F,1,H,W] float depth.
processors/mps_traj_processor.py
Reads closed-loop trajectories, optionally interpolates, outputs ts_world_device and timing; keeps gravity vector per sequence.
processors/mps_semidense_processor.py
Loads semidense global points and per-frame observations; computes volume bounds and preserves per-point std/1/std. Key helper _find_matching_timestamps_in_df aligns to requested timestamps.
processors/mps_online_calib_processor.py
Optional time-varying intrinsics/extrinsics; returns projection_params[T,15] and ts_device_camera[T,C,3,4].
processors/obb2_gt_processor.py and processors/obb3_gt_processor.py
Parse ADT/ASE OBB GT (2D and 3D). The 3D processor recenters AABB, maps instance/category IDs via taxonomy, and outputs camera-specific boxes and world poses.
processors/efm_gt_processor.py
Wraps Obb3GtProcessor but returns nested per-camera OBB3 dicts compatible with EVL; useful for oracle RRI training.
sample_builders/atek_data_paths_provider.py
Resolves per-sequence file paths for ADT/ASE given data_root_path; supports both mesh+ATEK variants.
sample_builders/obb_sample_builder.py
Composes processors into AtekDataSample for CubeRCNN tasks; orchestrates timestamp subsampling.
sample_builders/efm_sample_builder.py
EFM-oriented builder that includes depth and OBB3; validates timestamp consistency between RGB, depth, and trajectory; raises if off by tolerance.
subsampling_lib/temporal_subsampler.py
Computes subsampling factor to hit target frequency; provides get_timestamps_by_sample_index used by builders and WDS writers.
data_loaders/atek_wds_dataloader.py
process_wds_sample (reshape, type cast, merge gt_data shards), select_and_remap_dict_keys, atek_default_collation_fn, and load_atek_wds_dataset/create_native_atek_dataloader.
data_loaders/cubercnn_model_adaptor.py and data_loaders/sam2_model_adaptor.py
Key maps + minimal transforms to feed CubeRCNN or SAM2; keep tensor dtypes and shapes aligned to those models’ expectations.
evaluation/surface_reconstruction/surface_reconstruction_metrics.py
evaluate_single_mesh_pair orchestrates mesh loading, optional gravity correction, surface sampling, and bidirectional point-to-mesh distances; thresholds default to 5 cm for AR tasks.
evaluation/static_object_detection/static_object_detection_metrics.py
AtekObb3Metrics wraps MeanAveragePrecision3D; supports per-class and per-volume-range metrics, and respects configurable max detections.
evaluation/static_object_detection/obb3_csv_io.py
CSV writers/readers for OBB3 eval; preserve timestamps and SE(3) fields for reproducibility.
viz/atek_visualizer.py and viz/cubercnn_visualizer.py
Matplotlib-based inspection of images, semidense points, OBBs, trajectories; used for quick QC before training/evaluation.

3 End-to-End Pipeline (build → flatten → WDS)

Dataclasses (external/ATEK/atek/data_preprocess/atek_data_sample.py):
- MultiFrameCameraData(images: Tensor["F,C,H,W"], capture_timestamps_ns: Tensor["F"], frame_ids: Tensor["F"], projection_params: Tensor[15], t_device_camera: Tensor["F,3,4"], camera_valid_radius: Tensor[1], …); methods to_flatten_dict() / from_flatten_dict.
- MpsTrajData(ts_world_device: Tensor["F,3,4"], capture_timestamps_ns: Tensor["F"], gravity_in_world: Tensor[3]).
- MpsSemiDensePointData(points_world: list[Tensor["Ni,3"]], dist_std: list[Tensor["Ni"]], inv_dist_std: list[Tensor["Ni"]], capture_timestamps_ns: Tensor["F"], points_volume_min/max: Tensor[3]).
- MpsOnlineCalibData(projection_params: Tensor["T,15"], ts_device_camera: Tensor["T,C,3,4"], capture_timestamps_ns: Tensor["T"]).
- AtekDataSample aggregates the above; to_flatten_dict() prefixes fields (mfcd#, mtd#, msdpd#, mocd#) and stores gt_data dict intact. Theory: flattening keeps shard-friendly keys while preserving tensor semantics for reconstruction and detection.
Processors (data_preprocess/processors/):
- aria_camera_processor.py – reads VRS streams; outputs MultiFrameCameraData. Theory: ensures RGB/SLAM frames align with rig poses.
- depth_image_processor.py – converts z-depth to distance; caches rays; outputs depth frames [F,1,H,W]. Theory: enables GT depth backprojection consistency.
- mps_traj_processor.py – closed-loop trajectory to ts_world_device. Theory: provides world poses for all modalities.
- mps_semidense_processor.py – loads semidense points + stats per timestamp. Theory: preserves uncertainty for later weighting.
- mps_online_calib_processor.py – time-varying intrinsics/extrinsics. Theory: avoids calibration drift over long snippets.
- obb2_gt_processor.py, obb3_gt_processor.py, efm_gt_processor.py – assemble 2D/3D OBB GT per timestamp/camera. Theory: maintains consistent category/instance mapping for EVL/OBB metrics.
Sample builders (data_preprocess/sample_builders/):
- AtekDataPathsProvider(data_root_path) → dict of raw file paths for ADT/ASE.
- ObbSampleBuilder(conf, vrs_file, sequence_name, mps_files, gt_files) → AtekDataSample with OBB GT.
- EfmSampleBuilder(conf, vrs_file, depth_vrs_file, sequence_name, mps_files, gt_files) → AtekDataSample tailored for EVL (includes depth + OBB3).
- CameraTemporalSubsampler(vrs_file, conf) → subsampled timestamp lists via get_timestamps_by_sample_index(i). Theory: controls snippet length/stride; guarantees even spacing.
Writing: AtekWdsWriter(output_tar)
- add_sample(sample: dict) flattens and writes; get_num_samples(), close(). Theory: deterministic flattening for WebDataset shards.

4 Flattened Schema Highlights (per snippet, unbatched)

Images: mfcd#camera-rgb+images [F,3,H,W] uint8; SLAM streams [F,1,H',W'].
Camera meta: mfcd#camera-<lbl>+projection_params [15] float32; +camera_valid_radius [1]; +t_device_camera [F,3,4].
Trajectory: mtd#ts_world_device [F,3,4] float32; mtd#capture_timestamps_ns [F]; mtd#gravity_in_world [3].
Semi-dense: msdpd#points_world stacked tensor with points_world_lengths to recover per-frame lists; points_volume_min/max [3]; per-point dist_std, inv_dist_std.
Online calib (optional): mocd#projection_params [T,15]; mocd#ts_device_camera [T,C,3,4].
gt_data: nested dict for obb2_gt, obb3_gt, scores, etc.; tensors stored as gt_data#<name>.pth.
Naming rule: <producer>#<cam>+<field> stays filename-safe inside WDS shards.

5 Loading & Adaptors

process_wds_sample(sample: dict) -> dict (data_loaders/atek_wds_dataloader.py)
Restacks JPEGs to tensors, merges gt_data#* back, unpacks semidense lists using stored lengths. Theory: lossless inverse of flattening.
select_and_remap_dict_keys(sample_dict, key_mapping) -> dict
Keeps/renames fields for downstream schemas (CubeRCNN, SAM2, EVL).
atek_default_collation_fn(samples: list[dict]) -> dict
Stacks tensors when shapes agree, leaves lists otherwise. Theory: mixed dtypes/shapes preserved for EVL padding later.
load_atek_wds_dataset(urls, nodesplitter=None, dict_key_mapping=None, data_transform_fn=None, batch_size=None, collation_fn=None, repeat_flag=False, shuffle_flag=False)
Returns iterable WebDataset pipeline.
create_native_atek_dataloader(..., num_workers)
Torch DataLoader wrapper for the above.
Model adaptors:
- CubeRCNNModelAdaptor.get_dict_key_mapping_all() – ATEK→CubeRCNN.
- Sam2ModelAdaptor.get_dict_key_mapping_all() – ATEK→SAM2.
- For EVL use EfmModelAdaptor (see efm3d_implementation.qmd).
Raw VRS helper: atek_raw_dataloader_as_cubercnn.py::AtekRawDataloaderAsCubercnn mirrors the interface without WDS.

6 Evaluation Modules

Surface reconstruction: evaluate_single_mesh_pair(pred_mesh_filename, gt_mesh_filename, sample_num=10000, step=50000, threshold=0.05, correct_mesh_gravity=False, cut_height=None, rnd_seed=42) (evaluation/surface_reconstruction/surface_reconstruction_metrics.py)
Returns (metrics, accuracy, completeness) where accuracy = pred→GT distances, completeness = GT→pred. Theory: bidirectional point-to-mesh distance approximates Chamfer; prec/recall/F@5 cm derived from distance histograms.
Helper: compute_pts_to_mesh_dist(pts, faces, vertices, step)
Projects points onto triangles (barycentric test) else nearest vertex; batch-processed for memory safety.
Static object detection: AtekObb3Metrics(class_metrics=False, max_detection_thresholds=100, ret_all_prec_rec=False) (evaluation/static_object_detection/static_object_detection_metrics.py)
Methods: update(pred_dict, tgt_dict), compute(), reset(). Theory: wraps MeanAveragePrecision3D for ATEK-format OBB dicts; IoU over oriented boxes.
IoU helper: eval_obb3_metrics_utils.IouOutputs(vol, iou).
CSV I/O:
- AtekObb3CsvWriter(output_filename) → write_from_atek_dict(atek_dict, score, timestamp_ns, flush_at_end)
- GroupAtekObb3CsvWriter(output_folder, output_filename) → per-sequence CSVs
- AtekObb3CsvReader(input_filename) → read_as_obb_dict()
  Theory: reproducible export/import for metric computation and viz.

7 Visualisation & Debug

NativeAtekSampleVisualizer (viz/atek_visualizer.py)
plot_atek_sample(sample_dict, show_gt=True, save_path=None); supports plotting images, semidense points, trajectories, OBB2/OBB3, and 3D overlays. Theory: quick sanity checks on flattened→loaded fidelity.
CubercnnVisualizer (viz/cubercnn_visualizer.py)
plot_cubercnn_dict(cubercnn_dict, save_path=None).

8 NBV / Oracle Tips

Coordinate convention: Aria world x left, y up, z forward. Camera pose: T_world_cam = T_world_device @ T_device_camera; depth points already in world frame.
Preserve sequence_name, snippet_id, and points_volume_min/max to pair shards with meshes (scene_ply_<scene>.ply).
Prefer load_atek_wds_dataset() + EfmModelAdaptor so EVL/EFM utilities see consistent padded tensors.

9 Underused / To Integrate Better

depth_image_processor.py – enable RGB-depth ingestion to compare candidate renderings with GT distance maps directly.
mps_online_calib_processor.py – incorporate time-varying intrinsics/extrinsics for long snippets to curb drift in RRI.
efm_gt_processor.py – use nested multi-camera OBB3 GT instead of per-camera flattening for better EVL alignment.
AtekObb3Metrics – run during candidate evaluation to measure view impact on OBB recall/precision.
NativeAtekSampleVisualizer – integrate into debug loops before RRI runs to catch schema/remap issues early.

--- title: "ATEK Implementation Index" format: html --- # Purpose ATEK is the production data pipeline for Aria Synthetic Environments. This index lists signatures, shapes, and theory for building/flattening/loading snippets and for evaluation utilities we depend on (plus items we underuse). ## Wikipedia theory primer - **SLAM (Simultaneous Localization and Mapping)**: joint estimation of a sensor’s pose and a map of the environment from the same observations—classically solved with probabilistic filters combining odometry and landmark updates. ATEK’s trajectory + semidense streams are the outputs of this SLAM stage. Source: [Wikipedia — Simultaneous localization and mapping](https://en.wikipedia.org/wiki/Simultaneous_localization_and_mapping). - **SE(3) pose representation**: Special Euclidean group combining 3D rotations (SO(3)) and translations; pose composition is associative and invertible, which keeps `T_world_device @ T_device_camera` well-defined for every frame. Sources: [Special Euclidean group](https://en.wikipedia.org/wiki/Special_Euclidean_group) and [Special orthogonal group](https://en.wikipedia.org/wiki/Special_orthogonal_group). # File-by-file implementation guide (ATEK) - `data_preprocess/atek_data_sample.py` Nested dataclasses for camera, trajectory, semidense, online calib, and GT containers; `to_flatten_dict`/`create_atek_data_sample_from_flatten_dict` handle stable WDS key naming and reconstruction. - `processors/aria_camera_processor.py` Reads VRS streams, extracts per-frame calibration, applies pixel transforms, and outputs `MultiFrameCameraData`; supports label-dependent resize/crop. - `processors/depth_image_processor.py` Loads depth VRS, converts z-depth to distance, precomputes pixel-to-ray cache, and aligns timestamps to RGB/trajectory; returns `[F,1,H,W]` float depth. - `processors/mps_traj_processor.py` Reads closed-loop trajectories, optionally interpolates, outputs `ts_world_device` and timing; keeps gravity vector per sequence. - `processors/mps_semidense_processor.py` Loads semidense global points and per-frame observations; computes volume bounds and preserves per-point std/1/std. Key helper `_find_matching_timestamps_in_df` aligns to requested timestamps. - `processors/mps_online_calib_processor.py` Optional time-varying intrinsics/extrinsics; returns `projection_params[T,15]` and `ts_device_camera[T,C,3,4]`. - `processors/obb2_gt_processor.py` and `processors/obb3_gt_processor.py` Parse ADT/ASE OBB GT (2D and 3D). The 3D processor recenters AABB, maps instance/category IDs via taxonomy, and outputs camera-specific boxes and world poses. - `processors/efm_gt_processor.py` Wraps `Obb3GtProcessor` but returns nested per-camera OBB3 dicts compatible with EVL; useful for oracle RRI training. - `sample_builders/atek_data_paths_provider.py` Resolves per-sequence file paths for ADT/ASE given `data_root_path`; supports both mesh+ATEK variants. - `sample_builders/obb_sample_builder.py` Composes processors into `AtekDataSample` for CubeRCNN tasks; orchestrates timestamp subsampling. - `sample_builders/efm_sample_builder.py` EFM-oriented builder that includes depth and OBB3; validates timestamp consistency between RGB, depth, and trajectory; raises if off by tolerance. - `subsampling_lib/temporal_subsampler.py` Computes subsampling factor to hit target frequency; provides `get_timestamps_by_sample_index` used by builders and WDS writers. - `data_loaders/atek_wds_dataloader.py` `process_wds_sample` (reshape, type cast, merge `gt_data` shards), `select_and_remap_dict_keys`, `atek_default_collation_fn`, and `load_atek_wds_dataset`/`create_native_atek_dataloader`. - `data_loaders/cubercnn_model_adaptor.py` and `data_loaders/sam2_model_adaptor.py` Key maps + minimal transforms to feed CubeRCNN or SAM2; keep tensor dtypes and shapes aligned to those models’ expectations. - `evaluation/surface_reconstruction/surface_reconstruction_metrics.py` `evaluate_single_mesh_pair` orchestrates mesh loading, optional gravity correction, surface sampling, and bidirectional point-to-mesh distances; thresholds default to 5 cm for AR tasks. - `evaluation/static_object_detection/static_object_detection_metrics.py` `AtekObb3Metrics` wraps `MeanAveragePrecision3D`; supports per-class and per-volume-range metrics, and respects configurable max detections. - `evaluation/static_object_detection/obb3_csv_io.py` CSV writers/readers for OBB3 eval; preserve timestamps and SE(3) fields for reproducibility. - `viz/atek_visualizer.py` and `viz/cubercnn_visualizer.py` Matplotlib-based inspection of images, semidense points, OBBs, trajectories; used for quick QC before training/evaluation. # End-to-End Pipeline (build → flatten → WDS) - Dataclasses (`external/ATEK/atek/data_preprocess/atek_data_sample.py`): - `MultiFrameCameraData(images: Tensor["F,C,H,W"], capture_timestamps_ns: Tensor["F"], frame_ids: Tensor["F"], projection_params: Tensor[15], t_device_camera: Tensor["F,3,4"], camera_valid_radius: Tensor[1], …)`; methods `to_flatten_dict()` / `from_flatten_dict`. - `MpsTrajData(ts_world_device: Tensor["F,3,4"], capture_timestamps_ns: Tensor["F"], gravity_in_world: Tensor[3])`. - `MpsSemiDensePointData(points_world: list[Tensor["Ni,3"]], dist_std: list[Tensor["Ni"]], inv_dist_std: list[Tensor["Ni"]], capture_timestamps_ns: Tensor["F"], points_volume_min/max: Tensor[3])`. - `MpsOnlineCalibData(projection_params: Tensor["T,15"], ts_device_camera: Tensor["T,C,3,4"], capture_timestamps_ns: Tensor["T"])`. - `AtekDataSample` aggregates the above; `to_flatten_dict()` prefixes fields (`mfcd#`, `mtd#`, `msdpd#`, `mocd#`) and stores `gt_data` dict intact. **Theory**: flattening keeps shard-friendly keys while preserving tensor semantics for reconstruction and detection. - Processors (`data_preprocess/processors/`): - `aria_camera_processor.py` – reads VRS streams; outputs `MultiFrameCameraData`. Theory: ensures RGB/SLAM frames align with rig poses. - `depth_image_processor.py` – converts z-depth to distance; caches rays; outputs depth frames `[F,1,H,W]`. Theory: enables GT depth backprojection consistency. - `mps_traj_processor.py` – closed-loop trajectory to `ts_world_device`. Theory: provides world poses for all modalities. - `mps_semidense_processor.py` – loads semidense points + stats per timestamp. Theory: preserves uncertainty for later weighting. - `mps_online_calib_processor.py` – time-varying intrinsics/extrinsics. Theory: avoids calibration drift over long snippets. - `obb2_gt_processor.py`, `obb3_gt_processor.py`, `efm_gt_processor.py` – assemble 2D/3D OBB GT per timestamp/camera. Theory: maintains consistent category/instance mapping for EVL/OBB metrics. - Sample builders (`data_preprocess/sample_builders/`): - `AtekDataPathsProvider(data_root_path)` → dict of raw file paths for ADT/ASE. - `ObbSampleBuilder(conf, vrs_file, sequence_name, mps_files, gt_files)` → `AtekDataSample` with OBB GT. - `EfmSampleBuilder(conf, vrs_file, depth_vrs_file, sequence_name, mps_files, gt_files)` → `AtekDataSample` tailored for EVL (includes depth + OBB3). - `CameraTemporalSubsampler(vrs_file, conf)` → subsampled timestamp lists via `get_timestamps_by_sample_index(i)`. **Theory**: controls snippet length/stride; guarantees even spacing. - Writing: `AtekWdsWriter(output_tar)` - `add_sample(sample: dict)` flattens and writes; `get_num_samples()`, `close()`. **Theory**: deterministic flattening for WebDataset shards. # Flattened Schema Highlights (per snippet, unbatched) - Images: `mfcd#camera-rgb+images [F,3,H,W] uint8`; SLAM streams `[F,1,H',W']`. - Camera meta: `mfcd#camera-<lbl>+projection_params [15] float32`; `+camera_valid_radius [1]`; `+t_device_camera [F,3,4]`. - Trajectory: `mtd#ts_world_device [F,3,4] float32`; `mtd#capture_timestamps_ns [F]`; `mtd#gravity_in_world [3]`. - Semi-dense: `msdpd#points_world` stacked tensor with `points_world_lengths` to recover per-frame lists; `points_volume_min/max [3]`; per-point `dist_std`, `inv_dist_std`. - Online calib (optional): `mocd#projection_params [T,15]`; `mocd#ts_device_camera [T,C,3,4]`. - `gt_data`: nested dict for `obb2_gt`, `obb3_gt`, `scores`, etc.; tensors stored as `gt_data#<name>.pth`. - Naming rule: `<producer>#<cam>+<field>` stays filename-safe inside WDS shards. # Loading & Adaptors - `process_wds_sample(sample: dict) -> dict` (`data_loaders/atek_wds_dataloader.py`) Restacks JPEGs to tensors, merges `gt_data#*` back, unpacks semidense lists using stored lengths. **Theory**: lossless inverse of flattening. - `select_and_remap_dict_keys(sample_dict, key_mapping) -> dict` Keeps/renames fields for downstream schemas (CubeRCNN, SAM2, EVL). - `atek_default_collation_fn(samples: list[dict]) -> dict` Stacks tensors when shapes agree, leaves lists otherwise. **Theory**: mixed dtypes/shapes preserved for EVL padding later. - `load_atek_wds_dataset(urls, nodesplitter=None, dict_key_mapping=None, data_transform_fn=None, batch_size=None, collation_fn=None, repeat_flag=False, shuffle_flag=False)` Returns iterable WebDataset pipeline. - `create_native_atek_dataloader(..., num_workers)` Torch DataLoader wrapper for the above. - Model adaptors: - `CubeRCNNModelAdaptor.get_dict_key_mapping_all()` – ATEK→CubeRCNN. - `Sam2ModelAdaptor.get_dict_key_mapping_all()` – ATEK→SAM2. - For EVL use `EfmModelAdaptor` (see `efm3d_implementation.qmd`). - Raw VRS helper: `atek_raw_dataloader_as_cubercnn.py::AtekRawDataloaderAsCubercnn` mirrors the interface without WDS. # Evaluation Modules - Surface reconstruction: `evaluate_single_mesh_pair(pred_mesh_filename, gt_mesh_filename, sample_num=10000, step=50000, threshold=0.05, correct_mesh_gravity=False, cut_height=None, rnd_seed=42)` (`evaluation/surface_reconstruction/surface_reconstruction_metrics.py`) **Returns** `(metrics, accuracy, completeness)` where accuracy = pred→GT distances, completeness = GT→pred. **Theory**: bidirectional point-to-mesh distance approximates Chamfer; prec/recall/F@5 cm derived from distance histograms. - Helper: `compute_pts_to_mesh_dist(pts, faces, vertices, step)` Projects points onto triangles (barycentric test) else nearest vertex; batch-processed for memory safety. - Static object detection: `AtekObb3Metrics(class_metrics=False, max_detection_thresholds=100, ret_all_prec_rec=False)` (`evaluation/static_object_detection/static_object_detection_metrics.py`) Methods: `update(pred_dict, tgt_dict)`, `compute()`, `reset()`. **Theory**: wraps `MeanAveragePrecision3D` for ATEK-format OBB dicts; IoU over oriented boxes. - IoU helper: `eval_obb3_metrics_utils.IouOutputs(vol, iou)`. - CSV I/O: - `AtekObb3CsvWriter(output_filename)` → `write_from_atek_dict(atek_dict, score, timestamp_ns, flush_at_end)` - `GroupAtekObb3CsvWriter(output_folder, output_filename)` → per-sequence CSVs - `AtekObb3CsvReader(input_filename)` → `read_as_obb_dict()` **Theory**: reproducible export/import for metric computation and viz. # Visualisation & Debug - `NativeAtekSampleVisualizer` (`viz/atek_visualizer.py`) `plot_atek_sample(sample_dict, show_gt=True, save_path=None)`; supports plotting images, semidense points, trajectories, OBB2/OBB3, and 3D overlays. **Theory**: quick sanity checks on flattened→loaded fidelity. - `CubercnnVisualizer` (`viz/cubercnn_visualizer.py`) `plot_cubercnn_dict(cubercnn_dict, save_path=None)`. # NBV / Oracle Tips - Coordinate convention: Aria world x left, y up, z forward. Camera pose: `T_world_cam = T_world_device @ T_device_camera`; depth points already in world frame. - Preserve `sequence_name`, `snippet_id`, and `points_volume_min/max` to pair shards with meshes (`scene_ply_<scene>.ply`). - Prefer `load_atek_wds_dataset()` + `EfmModelAdaptor` so EVL/EFM utilities see consistent padded tensors. # Underused / To Integrate Better - `depth_image_processor.py` – enable RGB-depth ingestion to compare candidate renderings with GT distance maps directly. - `mps_online_calib_processor.py` – incorporate time-varying intrinsics/extrinsics for long snippets to curb drift in RRI. - `efm_gt_processor.py` – use nested multi-camera OBB3 GT instead of per-camera flattening for better EVL alignment. - `AtekObb3Metrics` – run during candidate evaluation to measure view impact on OBB recall/precision. - `NativeAtekSampleVisualizer` – integrate into debug loops before RRI runs to catch schema/remap issues early.