Title: DRAFT – Recurrent field transformations by distilled pairs for optical flow
Abstract
This presentation addresses the challenge of using learning-based algorithms for 3D scene reconstruction on resource-constrained end-user devices. Although integrating deep learning methods into the reconstruction pipeline has demonstrated superior performance to conventional techniques, the resulting large models may be impractical for resource-constrained devices. We propose an effective solution by introducing a method for compressing the deep learning models used in 3D reconstruction workflows. Our approach, called DRAFT, uses knowledge distillation (KD), adapted and extended to the complex feature and context extraction tasks related to optical flow. New distillation components based on algebraic sign pattern matrices (SPM) and inertia have been introduced to improve the KD process.
Speaker
Yanick Christian TCHENKO, PhD student IBISC, IRA2 team, supervised by Hedi TABIA (PR IBISC, IRA2 team) and Hicham HADJ-ABDELKADER (MCF IBISC, SIAM team)
NOTA
IRA2 seminars are held on the first Monday of each month, at the initiative of the IRA2 team management.
- Date: 08/04/2024, 1:30 pm
- Location: IBISC, Pelvoux site, room Ax101, see the seminar on Zoom: https://univ-evry-fr.zoom.us/j/95109429789?pwd=TEFjUERINWhpdnFhNDRvYmtkbm8yQT09
- Organized by: Hédi TABIA (PR Univ. Évry, IBISC IRA2 team)
