To put it simply, ShareTheBoard is an app that allows you to use a whiteboard with colleagues and friends, regardless of their location. It transforms your written ideas into crystal-clear digital content and automatically saves your work as you go.

It was designed to fill out the hole in the current market – many digital whiteboards usually focus solely on sharing the content on the board without capturing the person writing it. On the other hand, relying only on conference tools and merely pointing your laptop camera at a whiteboard usually results in an image that is too small, illegible, and often obstructed.

Basically, ShareTheBoard brings whiteboards into an otherwise predominantly digital world by delivering a more effective, engaging, and human-centric collaboration experience.

So far ShareTheBoard was recognized as:

• a winner of District Administration’s Top Ed Tech Products Awards at the Future of Education Technology® Conference (FETC)
• one of the Top 10 Distance and Remote Learning Solution Companies of 2022 by the publication Education Technology Insights (ETI)

To bring ShareTheBoard’s concept to life, the team needed to create an algorithm capable of doing real-time semantic segmentation. It had to classify what objects on the whiteboard should be considered content, what belonged to the background, and what could be a potential disruption.

Additionally, ShareTheBoard’s neural network had to work efficiently on high-resolution images while being small and efficient enough to accommodate not only older computers but also run smoothly through a web browser.

To sum it up, our team had to bridge the gap between two seemingly different worlds—working on less powerful processors while delivering high-definition results.

We started by researching already existing fully convolutional neural networks like DeepLabv3, but unfortunately, they were too slow to operate efficiently within web browsers.

Our team then focused on developing our own almost fully convolutional network, testing thousands of experimental models through a grid search for hyperparameters to identify the best one.

A challenge unique to this project arose in pixel classification, especially when dealing with shades of gray on a whiteboard. It was hard to judge objectively – training a neural network on a broad range led to disruptions appearing as content. On the other hand, a narrow range resulted in the actual content appearing as a dashed line.

To tackle this, we prepared A/B tests - we asked our testers to assess the performance of different algorithms. Following that, we created a metric that could objectively estimate output quality, aligning with the judgments made by people during A/B tests. We then used this metric to automate the evaluation of other models.

Next, we created our own data set to train and test how well neural network algorithms can correctly classify objects. To do this, we used:

• films and images created by our research team featuring different environments and backgrounds,
• a small publicly available data set,
• videos obtained from users of the prototype application,
• publicly shared videos featuring creating content on a whiteboard.

That is what led to the development of our neural network called “STB-745.” It excels in recognizing disruptions caused by human presence and effectively identifies content on the whiteboard, ensuring smooth real-time collaboration without the need for dedicated hardware or extensive training.

• We developed from scratch a neural network in less than 12 months
• The app runs in a web browser, making it compatible with most laptops and eliminating the need for native installations, allowing ShareTheBoard to acquire a wider audience
• Our STB-745 neural network achieves an impressive 96.3% accuracy in content identification, making using the app reliable and hassle-free
• On different devices the algorithm identifies content in less than 1 second, ensuring a smooth user experience