How Fitatom Reduced Labeling Time
From Two Weeks to Two Days
To create their algorithms, the team turned to human-in-the-loop (HITL) services and in-house tools to label large, diverse datasets for human biometrics. Without effective data and project management, throughput was too low, and overhead too high.
Superb AI’s image annotation tool for keypoints, alongside manual review processes and advanced data and user management.
86% reduction in labeling and model training time
3x increase in weekly labeling throughput and capacity
10% improvement in keypoint model accuracy
Founded in 2020, FitAtom emerged to fill a market void for harnessing computer vision to prevent avoidable injuries and accelerate rehabilitation. Boasting real-time motion tracking, an extensive workout library, and the capacity to track over 30 human body keypoints without biomarkers or sensors, FitAtom's technology analyzes posture and delivers biomechanical insights for a deeper understanding of client performance.
Before discovering Superb AI, FitAtom's team faced difficulties expanding their models due to time-consuming manual labeling and data management bottlenecks. Outsourcing and adopting human-in-the-loop (HITL) workflows fell short of expectations, while various commercial and open-source tools failed to meet their keypoint annotation requirements. Developing an in-house labeling tool based on the Django framework also proved burdensome. These difficulties resulted in 2-3 week gaps in product development, slowing their go-to-market efforts.
FitAtom adopted the Superb AI platform and image annotation tools to tackle these issues. This gave the team a centralized interface for precise keypoint annotation, comprehensive data asset management, and streamlined labeling project management. By integrating human-in-the-loop services within the platform and incorporating it into existing workflows, FitAtom's team significantly expedited labeling and iteration, substantially reducing the time required to enhance existing models or train new ones.
FitAtom also fine-tuned its data collection process more effectively by combining platform-provided metrics and in-house measurements to determine model performance on various images. They then used this data to establish new ground truth training datasets for further product development.
As a result, FitAtom experienced an immediate 86% reduction in labeling time and a 10% improvement in model performance. This accelerated product-market fit evaluation with clients, allowing them to transition from monthly to bi-weekly sprints, as building and validating new keypoint-based models now took 2-3 days rather than 2-3 weeks. The time and monetary savings from enhanced annotation and data management efficiency are now being redirected to further develop the company’s mobile app and other innovative capabilities.