Glossary

1 Glossary

1.1 Abbreviations and Terms

ASE - Aria Synthetic Environments: A large-scale dataset of 100K unique procedurally-generated indoor scenes with simulated Aria glasses sensor characteristics, egocentric trajectories, and ground truth annotations.
CPF - Central Pupil Frame: The coordinate frame placed at the midpoint between the eye boxes of the left and right eye of Aria glasses. Used as the reference frame for eye gaze vectors.
EFM3D - Egocentric Foundation Model in 3D: A foundation model that operates on various egocentric modalities (RGB, grayscale, semi-dense points).
EVL - Egocentric Voxel Lifting: Architecture proposed in the EFM3D paper that lifts multi-view egocentric observations into a gravity-aligned 3D voxel feature volume.
GT - Ground Truth: Reference data or annotations that represent the true or expected values, used for training and evaluation.
Frustum: The three-dimensional region obtained by slicing a pyramid or cone with two parallel planes and discarding the outer caps. In computer graphics a viewing frustum encodes the camera’s visible volume: the near and far clipping planes bound the depth range, while the lateral planes are shaped by field-of-view and aspect ratio, producing a truncated pyramid that defines which objects project onto the image sensor [1].
MPS - Machine Perception Services: A suite of derived data processing services provided by Meta for Project Aria data, including SLAM, eye gaze tracking, hand tracking, and more.
NBV - Next-Best-View: The problem of autonomously selecting the optimal next camera viewpoint to maximize information gain during active 3D reconstruction.
OOB - Oriented Bounding Box: A 3D bounding box that can be arbitrarily rotated in space to tightly fit objects, unlike axis-aligned bounding boxes.
PC - Point Cloud: A set of 3D points representing the geometry of a scene or object, typically acquired through depth sensors or reconstruction algorithms.
RRI - Relative Reconstruction Improvement: A metric quantifying how much a candidate view would improve the quality of 3D reconstruction, accounting for coverage, novelty, and information gain.
SLAM - Simultaneous Localization and Mapping: A technique for building a map of an unknown environment while simultaneously tracking the position of the sensor within it.
Track: Temporal sequence of 2D detections of corresponding image features across successive frames. I.e. points with high-gradient pixels, with per-frame (u,v) coordinates, timestamps and associated camera ID. From that track, one corresponding 3D point can be filtered (i.e. using bundle adjustment) and parametrized with an inverse distance (uncertainty measure). Can be expressed as $\mathcal{T}={(u_k,v_k,\text{cam}*k,t_k)}*{k=0}^{N}$.
SSL - Structure Scene Language / SceneScript Language: A structured language format for representing 3D scene layouts using primitives like walls, doors, windows, and objects with geometric parameters.
MTD - Motion Trajectory Data: Device poses over time, typically represented as a sequence of 6-DoF transformations.
MFCD - Multi-Frame Camera Data: Synchronized camera streams from multiple cameras on the Aria glasses.
MSDPD - Multi-Semi-Dense Point Data: Semi-dense 3D point clouds generated by SLAM algorithms, providing sparse but informative depth information.
VIO - Visual-Inertial Odometry: A method of estimating pose and motion by combining visual features from cameras with inertial measurements from IMUs.
LUF - Left–Up–Forward: A camera coordinate convention where the X axis points left, the Y axis points up, and the Z axis points forward.
VRS - Virtual Reality Standard (file format): A file format developed by Meta for storing and accessing multi-modal sensor data streams efficiently, used for Project Aria recordings.
Snippet: A short, fixed-length temporal window (typically 1 s / 10 frames) of synchronized Aria sensor data (RGB, SLAM cameras, poses, semi-dense points) that EVL lifts into a gravity-aligned 3D voxel grid for inference.

References

[1]

Wikipedia contributors, “Frustum.” [Online]. Available: https://en.wikipedia.org/wiki/Frustum

--- title: "Glossary" format: html --- # Glossary ## Abbreviations and Terms **ASE** - Aria Synthetic Environments : A large-scale dataset of 100K unique procedurally-generated indoor scenes with simulated Aria glasses sensor characteristics, egocentric trajectories, and ground truth annotations. **CPF** - Central Pupil Frame : The coordinate frame placed at the midpoint between the eye boxes of the left and right eye of Aria glasses. Used as the reference frame for eye gaze vectors. **EFM3D** - Egocentric Foundation Model in 3D : A foundation model that operates on various egocentric modalities (RGB, grayscale, semi-dense points). **EVL** - Egocentric Voxel Lifting : Architecture proposed in the EFM3D paper that lifts multi-view egocentric observations into a gravity-aligned 3D voxel feature volume. **GT** - Ground Truth : Reference data or annotations that represent the true or expected values, used for training and evaluation. **Frustum** : The three-dimensional region obtained by slicing a pyramid or cone with two parallel planes and discarding the outer caps. In computer graphics a *viewing frustum* encodes the camera’s visible volume: the near and far clipping planes bound the depth range, while the lateral planes are shaped by field-of-view and aspect ratio, producing a truncated pyramid that defines which objects project onto the image sensor [@Frustum-Wikipedia-2025]. **MPS** - Machine Perception Services : A suite of derived data processing services provided by Meta for Project Aria data, including SLAM, eye gaze tracking, hand tracking, and more. **NBV** - Next-Best-View : The problem of autonomously selecting the optimal next camera viewpoint to maximize information gain during active 3D reconstruction. **OOB** - Oriented Bounding Box : A 3D bounding box that can be arbitrarily rotated in space to tightly fit objects, unlike axis-aligned bounding boxes. **PC** - Point Cloud : A set of 3D points representing the geometry of a scene or object, typically acquired through depth sensors or reconstruction algorithms. **RRI** - Relative Reconstruction Improvement : A metric quantifying how much a candidate view would improve the quality of 3D reconstruction, accounting for coverage, novelty, and information gain. **SLAM** - Simultaneous Localization and Mapping : A technique for building a map of an unknown environment while simultaneously tracking the position of the sensor within it. **Track** : Temporal sequence of 2D detections of corresponding image features across successive frames. I.e. points with high-gradient pixels, with per-frame (u,v) coordinates, timestamps and associated camera ID. From that track, one corresponding 3D point can be filtered (i.e. using bundle adjustment) and parametrized with an inverse distance (uncertainty measure). Can be expressed as $\mathcal{T}={(u_k,v_k,\text{cam}*k,t_k)}*{k=0}^{N}$. **SSL** - Structure Scene Language / SceneScript Language : A structured language format for representing 3D scene layouts using primitives like walls, doors, windows, and objects with geometric parameters. **MTD** - Motion Trajectory Data : Device poses over time, typically represented as a sequence of 6-DoF transformations. **MFCD** - Multi-Frame Camera Data : Synchronized camera streams from multiple cameras on the Aria glasses. **MSDPD** - Multi-Semi-Dense Point Data : Semi-dense 3D point clouds generated by SLAM algorithms, providing sparse but informative depth information. **VIO** - Visual-Inertial Odometry : A method of estimating pose and motion by combining visual features from cameras with inertial measurements from IMUs. **LUF** - Left–Up–Forward : A camera coordinate convention where the X axis points left, the Y axis points up, and the Z axis points forward. **VRS** - Virtual Reality Standard (file format) : A file format developed by Meta for storing and accessing multi-modal sensor data streams efficiently, used for Project Aria recordings. **Snippet** : A short, fixed-length temporal window (typically 1 s / 10 frames) of synchronized Aria sensor data (RGB, SLAM cameras, poses, semi-dense points) that EVL lifts into a gravity-aligned 3D voxel grid for inference.