Reading in the Wild - Columbus Subset

[📝 Blogpost] [📂 Project Page] [📊 Data Explorer] [📄 Paper]

Introduction

The Reading in the Wild dataset is the first-of-its-kind large-scale multimodal dataset collected using Meta's smart glasses under Project Aria. The dataset contains 100 hours of reading and non-reading egocentric videos in diverse and realistic scenarios. It also includes eye gaze and head pose data collected during reading and non-reading activities. The dataset can be used to develop models not only for identifying reading activities but also for classifying different types of reading actvities in real-world scenarios.

The dataset contains two subsets -- Seattle subset and Columbus subset. This repository is for the Columbus subset. The Seattle subset is maintained separately here.

Overview of Columbus Subset

The Columbus subset contains around 20 hours of data from 31 subjects containing reading and non-reading activities in indoor scenarios. It is collected with the objective for zero-shot experiments. It contains examples of hard negatives (where text is present but is not being read), searching/browsing (which gives confusing gaze patterns), and reading non-English texts (where reading direction differs).

As summarized in the following chart, the Columbus subset contains data collected from reading across three different medium types including digital, print, and objects. It also contains data collected from reading across three different types of contents, including paragraphs which have long continuous text, short texts such as posters and nutrition labels, and non-textual content such as illustrative diagrams.

Comparison to Existing Datasets

Compared to existing egocentric video datasets as well as reading datasets, our dataset is the first reading dataset that contains high-frequency eye-gaze, diverse and realistic egocentric videos, and hard negative (HN) samples.

Models

A base model (v1_default) trained on the training data of the Seattle subset can be found here. The model uses a 5° FoV RGB crop (64x64) from the RGB camera of the glasses centered on the wearer's eye gaze, 3D gaze velocities sampled at 60Hz spanning 2s from the eye tracking cameras and 3D head orientation and velocity sampled at 60Hz spanning 2s from the IMU sensors. The model can selectively work with any combination of these three modalities.

Besides the base model, the following variants are also provided here:

• v1_1s: uses a shorter 1s span for Gaze data
• v1_15Hz: uses a lower 15Hz sampling frequency for Gaze data
• v1_large: uses a larger RGB crop size of 128x128
• v1_medium: outputs categorical predictions for medium (no-read, 'print, 'digital and objects).
• v1_mode: outputs categorical predictions for reading modes(no-read, 'walk, 'out-loud, engaged, scan, write/type and skim).

For details of the base model and its variants, please refer to here.

Getting Started

Setup

Use conda to create a new environment and install the required packages. The codebase has been tested with Python 3.12 and PyTorch 2.4.

conda env create -f environment.yml

Once the environment is created, activate it:

conda activate ritw-osu

Download

Dataset

The dataset is hosted at HuggingFace Hub in this page. To download please run:

python -m ritw.download --config-name config.yaml

An example config file is provided in /config/download.yaml. The dataset will be downloaded to the folder indicated by local_dir. The config file allows you to specify filters to download a subset of the dataset.

Models

Download the models from here and put them inside the models/ folder.

Inference

The inference pipeline is configurable via a config file. An example config is shown below:

# Example: ../config/config.yaml
# conf/config.yaml
start_time: 0.0
snippet_gap: 0.01667  # roughly 1/60 seconds
mode: "folder" # folder/single. If mode is single, please provide input_filename in config. If mode is folder, inference will be done on all files in root dir.
modalities:
  - "gaze"
  - "imu"
  - "rgb"
  - "gaze+rgb"
  - "gaze+imu"
  - "imu+rgb"
  - "gaze+imu+rgb"
output_save_path: "output/"
root_dir: "/path/to/ritw/dataset/"
model_name:
  - "v1_default"
  - "v0"
num_workers: 4  # adjust based on available CPU cores

The config allows selecting modalities and models to infer on. Note that, to add new models, put the model in the ritw-osu/models directory and add the model name to the config file.

To run prediction, create a config.yaml file (also see predict.yaml for reference) and save to ritw-osu/config. Then use the following command:

python -m ritw.predict --config-name config.yaml

The command runs each file and model combinations in separate processes. The output is saved in the form of csv files in directory <output_save_path>/<model_name>.

Evaluation

The evaluation module allows you to benchmark the performance of the models based on the inference results. It utilizes metadata from the recordings, applies configurable filters to focus on a specific subset of the dataset, computes various performance metrics for each modality, and outputs a summary table in Markdown format.

Below is an example configuration file (conf/config.yaml) for the evaluation module:

# Example: ../config/config.yaml
metadata_file: "data/metadata.csv"
result_dir: "output/v1_default"
target_recall: 0.9
metrics:
  - "F1"
  - "Acc"
  - "P@R=0.9"
  - "T@R=0.9"
  - "Acc@R=0.9"
  - "F1@R=0.9"
  - "AUC"
modalities:
  - "gaze"
  - "rgb"
  - "imu"
  - "gaze+imu"
  - "imu+rgb"
  - "gaze+rgb"
  - "gaze+imu+rgb"
filters:
  ContainsNonText:
    - "images"
  ShortTextOrPara:
    - "paragraphs"
  Medium:
    - "digital"
  Platform:
    - "laptop"

After setting up your environment and ensuring that the metadata and prediction CSV files are available, run the evaluation module from the project’s root directory:

python -m ritw.evaluate --config-name config.yaml

License

Reading in the Wild - Columbus Subset dataset and code is released by The Ohio State University under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). Data and code may not be used for commercial purposes. For more information, please refer to the LICENSE file included in this repository.

Attribution

When using the dataset and code, please attribute it as follows:

@inproceedings{yang25reading,
      title={Reading Recognition in the Wild},
      author={Charig Yang and Samiul Alam and Shakhrul Iman Siam and Michael Proulx and Lambert Mathias and Kiran Somasundaram and Luis Pesqueira and James Fort and Sheroze Sheriffdeen and Omkar Parkhi and Carl Ren and Mi Zhang and Yuning Chai and Richard Newcombe and Hyo Jin Kim},
      booktitle={arXiv Preprint},
      year={2025},
      url={https://arxiv.org/abs/2505.24848},
}