Data for: The Bystander Affect Detection (BAD) Dataset for Failure Detection in HRI

<h3>Project Overview</h3>
For a robot to repair its own error, it must first know it has made a mistake. One way that people detect errors is from the implicit reactions from bystanders – their confusion, smirks, or giggles clue us in that something unexpected occurred. To enable robots to detect and act on bystander responses to task failures, we developed a novel method to elicit bystander responses to human and robot errors.</p>
<h3>Data Overview</h3>
This project introduces the Bystander Affect Detection (BAD) dataset – a dataset of videos of bystander reactions to videos of failures. This dataset includes 2,452 human reactions to failure, collected in contexts that approximate “in-the-wild” data collection – including natural variances in webcam quality, lighting, and background. The BAD dataset may be requested for use in related research projects. As the dataset contains facial video data of participants, access can be requested along with the presentation of a research protocol and data use agreement that protects participants.</p>
<h3>Data Collection Overview and Access Conditions</h3>
Using 46 different stimulus videos featuring a variety of human and machine task failures, we collected a total of 2,452 webcam videos of human reactions from 54 participants. Recruitment happened through the online behavioral research platform Prolific (https://www.prolific.co/about), where the options were selected to recruit a gender-balanced sample across all countries available. Participants had to use a laptop or desktop. Compensation was set at the Prolific rate of $12/hr, which came down to about $8 per participant for about 40 minutes of participation.</p>
Participants agreed that their data can be shared for future research projects and the data were approved to be shared publicly by IRB review. However, considering the fact that this is a machine-learning dataset containing identifiable crowdsourced human subjects data, the research team has decided that potential secondary users of the data must meet the following criteria for the access request to be granted:</p>
1. Agreement to three usage terms:</br>
- I will not redistribute the contents of the BAD Dataset</br>
- I will not use videos for purposes outside of human interaction research (broadly defined as any project that aims to study or develop improvements to human interactions with technology to result in a better user experience)</br>
- I will not use the videos to identify, defame, or otherwise negatively impact the health, welfare, employment or reputation of human participants</p>
2. A description of what you want to use the BAD dataset for, indicating any applicable human subjects protection measures that are in place. (For instance, "Me and my fellow researchers at University of X, lab of Y, will use the BAD dataset to train a model to detect when our Nao robot interrupts people at awkward times. The PI is Professor Z. Our protocol was approved under IRB #.")</p>
3. A copy of the IRB record or ethics approval document, confirming the research protocol and institutional approval.</p>
<h3>Data Analysis</h3>
To test the viability of the collected data, we used the Bystander Reaction Dataset as input to a deep-learning model, BADNet, to predict failure occurrence. We tested different data labeling methods and learned how they affect model performance, achieving precisions above 90%.</p>
<h3>Shared Data Organization</h3>
This data project consists of 54 zipped folders of recorded video data organized by participant, totaling 2,452 videos. The accompanying documentation includes a file containing the text of the consent form used for the research project, an inventory of the stimulus videos used, aggregate survey data, this data narrative, and an administrative readme file.</p>
<h3>Special Notes</h3>
The data were approved to be shared publicly by IRB review. However, considering the fact that this is a machine-learning dataset containing identifiable crowdsourced human subjects data, the research team has decided that potential secondary users of the data must meet specific criteria before they qualify for access. Please consult the Terms tab below for more details and follow the instructions there if interested in requesting access.