Collecting consented video datasets for training ML systems

The Challenge

The traditional approach involved numerous steps, including setting up cameras to record videos, labeling them, and sending these datasets to a verification team for quality control. This was cumbersome, expensive, biased, and most of the time did not get consent from subjects. The client wanted to rethink the existing dataset collection process as privacy-preserving by design.

The challenges involved were:

• Develop a solution that allows collectors worldwide to submit, edit, and label high-quality videos in real-time.
• Ensure videos contain only people who have consented to be included.
• Incorporate an easy and intuitive process of collecting, consenting, and verifying large datasets as simple as taking a selfie.
• Implement a secure method to safeguard curated video datasets.

The Solution

• BTC developed a solution comprising native Android and iOS mobile applications to collect large-scale, high-quality labeled datasets.
• The solution allowed users to easily record, annotate, and verify custom video content in real-time at low costs.
• Anonymized videos were collected in real-time by blurring out detected faces of non-consented subjects before uploading them.
• The platform included built-in pages to record universally-acceptable consent from subjects in written and audio formats.
• Enabled users to edit both objects and activities captured in the video and use multiple Bounding Boxes (rectangular borders enclosing an image) with animation to make changes. An offline editing option is also available for users.
• Provided gesture options such as tap, double tap, pinch for video editing, and other features to track frame coordinates of the video and handle video resolutions in different orientations.
• Users could also view, playback, and give screen-based feedback to videos uploaded by other users. This included activity-specific consensus questions for calculating the video’s recommendation percentage and generic feedback for analytics purposes.
• Designed a secure server-side storage environment to store curated videos and annotations in a structured manner, thereby protecting the privacy of video datasets.

Devices:

Android
iOS

Technology:

Java
Javascript
iQuery
Swift
UIKit
jQuery
AVFoundation
Xcode

The Outcome

• BTC’s solution helped the client build a massive library of consented and privacy-protected video datasets used for training machine learning systems.
• The platform also ensured that the consent of each subject included in the video is collected, and non-consented people are anonymized, preserving the privacy of individuals.
• Using our intelligent solution, the curation costs are an order of magnitude lower than existing traditional methods.
• People from over five continents and fifty countries are currently using the solution to collect real-life videos for global visual AI applications.