Our master thesis achieved 1st place in the faculty round of the Student Scientific Work competition in the section Data Analysis, Artificial Intelligence, and Informatics. You can view the presentation from the conference in Slovak at this link.

The master thesis deals with recognizing signs in sign language from videos. After analyzing multiple datasets, the Greek Sign Language (GSL) dataset containing 40,785 videos was chosen, each representing the corresponding sign to one of 310 possible words in Greek Sign Language.
For solving this problem, we also use hand skeleton representations generated using the Mediapipe library. Examples of skeletons for videos are shown in the images below:
Generated skeletons for video. Generated skeletons for video. Generated skeletons for video.
Generated skeletons for video.
To solve this task, we have employed a 3D CNN and we have achieved 83.16% accuracy during the validation phase after various experiments with the network architecture, as well as dataset preprocessing.
Example of the model output. Example of the model output. Example of the model output.
Example of the model output for specific inputs (from the validation set).