Scaling Up a Fitness App with Deep Neural Networks
Boosted application performance by 19%.
Boosted application performance by 19%.
- ChallengeOn-Demand Neural Network Development
- SolutionCNNs for Activity Recognition and Error Detection
- Technologies and toolsPyTorch, Keras, Tensorflow, CoreML, OpenCV, ONNX
Client
The Client is a fitness tech startup. They developed an app for at-home workouts. The main focus of the app is personalized fitness programs and zero equipment workouts. To make their app worth it, they wanted to empower it with AI.
The client has no in-house data science department. They were looking for experienced neural network developers. That was the reason they contacted the InData Labs team. They asked us to develop custom error-detection models and advance activity recognition ones.
Challenge: on-demand neural network development
The client already could recognize human activity in the app but was not satisfied with its quality. It was slow and inaccurate. So they asked us to advance the quality of activity recognition with deep learning.
In terms of the project, our neural network consultants suggested the following computer vision tasks:
| Object Localization and Object Detection | We offered to use Deep Learning algorithms for accurate and fast object detection. The main goal was to determine if there were any objects from the specified varieties like humans, animals, vehicles, etc.) and determine their boundaries. |
| Recognizing Human Motions and Activities | We decided to improve the human movement from sensor data with CNNs. Our tech experts supposed that they were best suited for learning data from a sensor (a smartphone accelerometer). |
| Error Detection | Our idea was to develop custom error detection algorithms. They would recognize errors and inaccuracies during workouts. If any error is detected, the AI coach will tell the user to exercise the right way. |
| Cross-Platform Development | The client had plans to run their app on various platforms. We offered our hands-on experience to make it run smoothly on different mobile devices. |
Solution: CNNs for activity recognition and error detection
To solve the project tasks, we did research. We studied the client’s challenges, the specifics of the app and the problems it had.
We set the following goals:
- Develop custom algorithms for object localization and detection
- Improve the quality of activity recognition
- Develop robust error detection algorithms
- Stabilize and optimize neural network architecture for the app to work fast and accurate
- Make the app run on various platforms (cross-platform development)
- Provide visualization tools on CNNs for easy decision making
- Provide scripts and guides for the client’s team for easy deployment on mobile devices
A pioneer in fitness and wellness project development, InData Labs had successfully done all the computer vision tasks related to the project. We improved the quality of activity recognition, created error detection algorithms from scratch. We also did other tasks that boosted app performance and speed.
Result: custom error detection algorithms, improvement in activity recognition
The InData Labs’ team has applied the latest technology to the client’s case. We did a lot of research and discussion to line up with the client’s requirements.
The benefits the client’s got from collaborating with us:
- activity recognition quality improved by 45%
- advanced object localization and detection
- false-positive detection decreased by 15%
- overall app speed improved by 30%
- bespoke error detection algorithms
- app performance boosted by 19%
- cross-platform development (neural networks can run on various mobile devices)
- neural networks can be converted to CoreML, TFLite, Tensorflow Mobile, MACE, etc.
As a result, the client’s got a solution catered specifically for their needs. Later on, they again contacted InData Labs for deep learning consulting and development.