23 set 2024
08:00 Master's Defense IC3 Auditorium
Theme
Gesture recognition in Brazilian sign language (Libras) using visual transformers
Student
Wladimir Arturo Garces Carrillo
Advisor / Teacher
Marcelo da Silva Reis - Co-supervisor: Emely Pujólli da Silva
Brief summary
The lack of accessibility for deaf communities is a persistent challenge. Despite advances in public policies and technology, there are still difficulties in communication and access to basic services for deaf people, who also face stigma and prejudice. One of the obstacles to the development of Automatic Sign Language Recognition (ASLR) technologies, especially in Brazilian Sign Language (Libras) is the scarcity of reliable databases, due to the lack of standardization terminologies, annotated samples of signs and regional variations. To mitigate limitations regarding the size of existing Libras data sets, in this work we propose the application of data augmentation techniques, based on the generation of videos with Probabilistic Diffusion Models for Noise Removal (DDPM). Probabilistic Models). To this end, we will use the Libras data set used by Vidalón and Martino [96] (Elias Dataset) for training and testing the learning model and Visual Transformers (ViT) as a learning model. For the augmentation process, we will also investigate the use of other Libras data sets existing in the literature. We hope that this work will enable the development of more accurate and efficient solutions for ASLR.
Examination Board
Headlines:
| Marcelo da Silva Reis | IC / UNICAMP |
| Ivani Rodrigues Silva | FCM / UNICAMP |
| Hélio Pedrini | IC / UNICAMP |
Substitutes:
| Andre Santanche | IC/UFBA |
| Kate Mamhy Oliveira Kumada | CCNH/UFABC |