The School of Computer Science is pleased to present…
Date: Monday, May 5th, 2025
Time: 12:00 PM
Location: Essex Hall, Room 122
Sign language (SL) serves as a vital mode of communication for millions of individuals worldwide, yet current SL avatar reconstruction systems face significant computational inefficiencies that hinder their deployment in real-time and resource-constrained environments. SGNify, a leading SL avatar synthesis framework, achieves high-fidelity reconstructions by processing every frame of a video input. However, this exhaustive approach results in considerable redundancy due to static or low-motion frames, leading to unnecessary computational overhead. This thesis introduces FESLAR (Frame-Efficient Sign Language Avatar Reconstruction). This novel optimization framework builds on SGNify by selectively processing only linguistically critical keyframes and employs interpolation techniques for frame reconstruction, maintaining semantic fidelity while significantly reducing processing time. FESLAR utilizes the FILM model for video frame generation and linear/RBF mesh interpolation for 3D avatar continuity. Quantitative evaluations across multiple SL video sequences demonstrate that FESLAR achieves up to 84% reduction in computational workload with minimal loss in visual quality, as measured by SSIM, RMSE, and mesh deformation metrics. This hybrid strategy balances efficiency and expressiveness, enabling scalable applications in education, live interpretation, and assistive technologies, and offers a significant advancement toward real-time, accessible sign language communication systems.
Internal Reader: Dr. Boubakeur Boufama
External Reader: Dr. Mohamed Belalia
Co-supervisor: Dr. Naimul Khan
Advisor: Dr. Imran Ahmad
Chair: Dr. Arunita Jaekel