eComm Image Product Labeling
and Validation Use Case

Summary

A Fortune 500 company with a large marketplace has recently partnered with Tasq.ai, an ultra-scale data labeling platform. The purpose of this partnership is the creation of high-quality datasets for training computer vision machine learning (ML) models to recognize products accurately, collect metadata, and better understand the intent of customers using the marketplace.

Tasq.ai accomplishes this task by using Tasqers – a diverse, unbiased global crowd of filtered human annotators. Tasqers have high cognitive abilities allowing Tasq.ai’s platform to manage the process of creating training datasets faster and with higher accuracy.

This process is carried out by using a codeless workflow data pipeline of nano-tasks. These nano-tasks are sent to Tasqers for human annotation and verification.

  1. Tasq.ai’s solution helps Data Science and ML teams to achieve high-quality, production-ready datasets for their artificial intelligence (AI) projects faster and economically than other solutions.
  2. Tasq.ai’s solution helped the F500 Data Science and ML team to achieve higher quality, production-ready datasets for their artificial intelligence (AI) projects faster and more economically than other solutions they tried before.

How does Tasq.ai provide its high-quality datasets and data labeling?

It does this by screening millions of online users in seconds to find those with the necessary cognitive capabilities to excel in a specific task. It then incentivizes those users to complete the task accordingly.

Tasq.ai achieves high-quality results with committed quality through the following methodology

Identified Tasqers
Identified Tasqers are tested against Ground Truth based tasks for qualification and accuracy purposes
Multiple Judgments
Multiple judgments are aggregated in order to reach pre-defined confidence scores.
Adaptive Sampling
Whether additional judgments are needed is determined through adaptive sampling.

Tasq.ai’s elastic platform enables the marketplace to increase or decrease its data labeling output within hours without delays.

These value propositions give the marketplaces a competitive edge in achieving throughput for creating high-quality datasets for ML and maintaining the quality levels throughout the process.

Approach

Tasq.ai’s Approach
  • Determine if an image/video has text in it, and if so, determine which language (using a model + human verification).
  • Route the image to the relevant demographic crowd for better understanding and localization.
  • Determine if a product has commercial intent or not.
  • Classify the products according to a predefined workflow of questions and label them accordingly with multiple judgments for high confidence score precisions.
  • Compare products from different catalogs and add attributes.

Outcome

  • Accuracy – After a few initial iterations, Tasq.ai was able to reach averages of above 95% accuracy per single question and overall 90-99% accuracy scores per entire workflow. This outcome was received by aggregating 3-7 human judgments for every question. A Dynamic Configuration of judgments was deployed (the collection of judgments until a specific threshold/confidence level is reached). The accuracy levels are expected to improve mostly due to additional quality measures, assisting models, and demographic targeting.
  • Scale & Elasticity – After automating the above process, Tasq.ai is now supplying the Marketplace team with hundreds of thousands of annotations each and every day and has the scalability to reach millions of annotations if requested with no effort and no need for long onboarding.
  • Unbiased Results – Another advantage of Tasq.ai’s platform is the unbiased distribution of the global Tasqers. Their platform also enables the ability to target specific demographics per request.

Screenshot from the Workflow on the Platform

Screenshot from the Workflow on the Platform
Screenshot from the Workflow on the Platform

demo-30min-book

Book Your 30-Minute Demo Today!

Contact us