By 
Fitness apps are becoming increasingly dependent on AI to power their offerings by providing AI-based workout analysis, which incorporates technologies such as computer vision, human pose estimation, and natural language processing techniques, as more and more people use fitness apps to train and track their development and performance.
Datagen is a fitness AI startup located in Tel Aviv that was established in 2018 and promises to produce “high-performance synthetic data, with an emphasis on data for human-centric computer vision applications.”
On the firm’s self-service, visual synthetic data platform, the company today made an announcement about a new domain called Smart Fitness. This new domain will assist AI developers in producing the data they need to evaluate individuals exercising and teach smart fitness equipment to “see.”
According to statements made by Ofir Zuk, CEO of Datagen, in an interview with VentureBeat, the company’s primary objective is to assist computer vision teams in accelerating the creation of human-centric computer vision jobs. Nearly every example of AI’s application that we’ve seen so far involves people. We are especially working toward a solution to, as well as a better understanding of, the connectivity that exists between persons and the settings in which they interact. We refer to it as being “human in context.”
The gym is represented by synthetic visual data.
With the Smart Fitness platform, you may access visual data in the form of 3D-annotated videos and still photographs. Useful for activities like body key point estimate, pose analysis, posture analysis, repetition counting, object recognition, and more, this visual data faithfully depicts fitness settings, high-level motion, and human-object interactions.
The technology also allows teams to rapidly iterate and enhance their model’s performance by generating data for the whole body in motion. The Smart Fitness platform, for instance, has the benefit of allowing users to rapidly simulate multiple camera types for gathering a wide range of diverse exercise synthetic data, which may be used in situations of posture estimation analysis.
Obstacles to overcome while trying to get AI in shape
One of the one-of-a-kind answers that artificial intelligence (AI) has to offer is something called pose estimation. Pose estimation is a computer vision approach that helps estimate the position and orientation of the human body using a picture of a person. It has a number of applications, including worker posture analysis and markerless motion capture, for instance, as well as the animation of avatars for use in artificial reality.
In order to conduct an accurate analysis of posture, it is important to take many pictures of the human actor in question while they are interacting with their surroundings. Next, a trained convolutional neural network is used to these pictures in order to make a prediction about the positions of the human actor’s joints inside the image. AI-based fitness applications often make use of the device’s camera, capturing videos at resolutions of up to 720p and 60 frames per second (fps), so that more frames may be captured during activity.
The challenge is that in order to train AI for fitness analysis using a method such as posture estimation, computer vision specialists require access to enormous volumes of visual data. The data that involves individuals carrying out exercises in a variety of formats and interacting with a number of different things is quite complicated. To eliminate the possibility of bias, the data must have a large variance and be sufficiently varied. It is quite difficult to collect reliable data that encompasses such a wide diversity of situations. In addition to these drawbacks, hand annotation is labor-intensive, time-consuming, and fraught with the possibility of making mistakes.
Although a satisfactory degree of accuracy in 2D pose estimation has already been achieved, there is still room for improvement in 3D pose estimation with regard to the generation of correct model data. This is particularly true when drawing conclusions from a single photograph and having no knowledge about the depth of the scene. Some approaches include using a number of cameras all aimed in the same direction at the subject, collecting data from several depth sensors in order to create more accurate predictions.
However, one of the challenges associated with 3D pose estimation is the dearth of big datasets of humans posing in open locations that have been annotated. Large datasets for 3D pose estimation, such as Human3.6M, were collected fully inside to reduce the amount of visual noise that was present throughout the process.
The creation of new datasets that include data on ambient circumstances, clothing diversity, powerful articulations, and other significant aspects is a continuous endeavor that is now being made.
Methods for Easily Displaying Exercises
With the help of Datagen’s smart fitness platform, businesses are able to generate tens of thousands of distinct identities in a fraction of the time it would normally take to do so. These identities carry out a range of exercises in a number of settings.
According to Zuk, “with the power of synthetic data, teams may produce all the data they want in a matter of a few hours with particular specifications.” This not only assists in retraining the network and machine learning model, but it also enables you to have it fine-tuned in a very short amount of time.
Use of Artificial Intelligence and Synthetic Data in the Workplace Fitness
It will see growing adoption for use cases where data must be guaranteed to be anonymous or privacy must be preserved (such as medical data), augmentation of real data, especially where data collection costs are high, where there is a need to balance class distribution within existing training data (such as with population data), and emerging AI use cases for which limited real data is available.
Datagen’s value proposition depends on many of these use cases. Increasing data quality, covering a broad range of circumstances, and preserving privacy during ML training can help smart fitness gadgets or applications, he added.
Zuk recognizes it’s early days for AI, synthetic data, and digital fitness technology.
“They’re non-reactive and lean,” he remarked. Adding visual features to fitness applications will enhance them, particularly as people workout at home more. We see rising demand and can only speculate what others will do with our data.
