One of the ways the Visual Intelligence team at Samsung R&D Institute Ukraine (SRK) makes sure to offer users the best possible experience is by taking part in relevant competitions. This allows the team to see how well their solutions stand against other researchers in the field.

In June, the Conference on Computer Vision and Pattern Recognition (CVPR-2021) took place virtually. Today, CVPR is the premier annual computer vision event and is regarded as one of the most important conferences in the field. The main topics of interest are related to the extraction of structures or answers from images or video, or the application of mathematical methods and deep learning to data for pattern recognition. The CVPR event consists of a main conference and several co-located workshops and competitions.

The Creative Input Intelligence team from SRK could not miss the chance to take part in the Chart Question Answering Challenge.

The challenge was held as part of the CVPR-2021 Chart Question Answering Workshop. The field of chart question answering is an emerging research area that requires image and text analysis combination. The goal is to not only locate and recognize charts, but also analyze element proportions and relative positions and correctly identify the characteristics of the required data points.

One of SRK's missions is to ensure seamless recognition of all types of documents in all contexts and environments. With this in mind, the Chart Question Answering Challenge was a perfect opportunity to break new ground and put SRK’s approach to document recognition to the test. 

The Chart Question Answering Challenge included three tasks: measuring Cleveland and McGill’s Angle and Length stimuli; comparing ratios in bar and pie charts; and answering textual questions regarding the charts. 

The team from SRK swept the board this year, leaving competitors behind in each of the three sessions. 

“It gets a little challenging with semantic questions. And the ultimate generalization test was training on levels of apples and then testing on levels of oranges, completely removing the content and just focusing on solving the task without knowing the domain before ... For level one and two, the results were measured with RMSE. For level three, we had a special error metric Entropy plus RMSE plus Levenshtein Edit Distance to compare the answers. And the team from SRK has the lowest errors' values for all three levels. Congratulations to the SRK team,” commented a member of the organizing committee during the workshop when announcing the winner.

This was not the only success this summer. SRK also demonstrated top-tier results in the Document Layout Analysis Task at the 16th International Conference on Document Analysis and Recognition (ICDAR-2021) competition on Scientific Literature Parsing. The competition specialized in analyzing the layout of scientific paper images. This focus is also very important, as a good performance in such tasks allows for an excellent user experience in educational and professional scenarios. 

The main objective of SRK’s Creative Input Intelligence team has always been to bring Samsung’s users high quality products that make text input and interaction easy and enjoyable. Check out these state-of-the-art text recognition technologies in the Samsung Notes* and CalliScan** apps available through the Galaxy Store. 

*Samsung’s Handwriting Recognition (HWR) Solution has proven its status as the world-best handwriting recognition solution and every year provides innovative features in the Samsung Notes app. 

**The CalliScan app converts handwritten notes to digital text.