Aria Gen 2 applications are open. Our next generation hardware unlocks new possibilities.

INTRODUCING

Reading Recognition in the Wild

100 hours of egocentric RGB video, eye gaze, and head pose data of reading and non-reading activities collected in diverse and realistic scenarios.

WHAT IS IT?

100 hours of reading recognition in-the-wild

A first-of-its-kind dataset, Reading Recognition in the Wild is designed to help solve the task of reading recognition by wearable devices in diverse environments. Reading Recognition in the Wild is a large-scale multimodal dataset comprising 100 hours of reading and non-reading videos captured in diverse and realistic scenarios using Aria Gen 1 glasses. The dataset features video, eye gaze, and head pose sensor outputs, created to help solve the task of reading recognition from a wearable device.

Dataset Contents

  • 100 hours of video data
  • 1716 sequences
  • 111 participants
  • 60 Hz high-frequency eye tracking
  • 19 high-level scenarios of reading and non-reading activities
  • 150+ Reading materials
  • Includes non-reading activities containing both daily activities and hard negatives (where text is present but is not being read).

Sensor Data

  • 1 x 110 degree FOV + 30 Hz Rolling Shutter RGB camera
  • 2 x 80 degree FOV + 60 Hz, Global Shutter mono cameras for eye-tracking with IR illumination
  • 2 x 150 degree FOV + 30 Hz Global Shutter mono cameras for SLAM and hand tracking
  • 1 kHz + 800 Hz IMUs + barometer & magnetometer environmental sensors
  • 7 x 48 KHz spatial microphones

Annotations

  • Start and end time of reading activities in each sequences
  • Type of Reading material used
  • Type of reading
  • The participant’s description of the reading material
  • 3D eye gaze
  • 6DoF device trajectory

WHAT MAKES IT SPECIAL?

Real-world multimodal egocentric data for reading recognition in the wild

Diverse scenes and scenarios

  • 19 high-level scenarios of reading and non-reading activities
  • Various indoor and outdoor locations: Offices, Libraries, Lounges, Homes, Stores, Balconies, Patios, Roads, Trails, In the Woods.

Various Reading Modes

Our dataset includes multiple reading modes including:

  • Engaged reading
  • Skimming
  • Scanning
  • Reading out loud
  • Multi-tasking (While walking, Writing/typing)

Comprehensive Reading Materials

150+ Reading materials in different mediums and text lengths.

Annotations

The recordings were processed by Aria Machine Perception Service to obtain accurate 6 DoF device trajectory, semi-dense point clouds, and 3D eye gaze ray estimation with depth.

In addition, the annotation contains:

  • Start and end time of reading activities in each sequences
  • Type of reading material used
  • Type of reading mode
  • The participant’s description of the reading material

Seattle and Columbus subsets

The Reading Recognition in the Wild dataset contains two distinct subsets, one captured in Columbus and the other captured in Seattle. The Seattle subset focuses on diversity, while the Columbus subset aims to test our model’s ability to generalize in unseen settings, as well as identify edge cases where the model fails.

Seattle subset

The Seattle subset emphasizes diversity, comprising 80 hours of data from 80 participants engaged in various reading and non-reading activities across multiple indoor and outdoor settings.

Columbus subset

The Columbus subset is designed to evaluate edge cases, comprising approximately 20 hours of data from 31 subjects containing reading and non-reading activities in indoor settings. It also features 4 different languages.