Thesis

My thesis focuses on enhancing variance reduction techniques to improve gradient quality within differentiable applications, specifically leveraging morphable models. Building upon the foundations set by the NextFace and Flame projects, this work aims to refine the computational approach for more accurate and stable outcomes in 3D facial reconstruction. The expected completion and delivery of this research are set for September 2024.

Flame face reconstruction

Next Face

Contribution to Ubisoft Laforge Research in Open Source Project:
Implemented a new inverse rendering engine for high-fidelity facial reconstruction, contributing to the NextFace project. This work is grounded in a series of papers by Abdallah Dib and others, advancing the field of face reconstruction with differentiable rendering techniques.

• GitHub Repository: NextFace
• Research Papers: Link to papers by Abdallah Dib
• Public Pull Request: GitHub Pull Request

Physics Based Renderer

Project Highlights:
This section outlines the key features implemented in the renderer and the advanced rendering concepts explored throughout the project.

Implemented Features:

• Oren-Nayar diffuse model
• Microfacet model for transmissive, conductive, and dielectric materials
• Geometric intersections (spheres, squares, and meshes)
• BVH structures for efficient scene traversal
• Surface Area Heuristic (SAH) for optimal BVH
• Texture sampling and Normal Maps
• Procedural textures for dynamic surface detailing
• Custom integrators for flexible rendering algorithms
• Multiple Importance Sampling (MIS) for balanced light sampling
• Custom camera models for diverse perspectives
• Fresnel Blend material for realistic surface interactions
• Lambertian, metallic, and dielectric materials
• Recursive ray-tracing for accurate light paths
• Volumetric rendering for atmospheric effects
• Environment Mapping with importance sampling for realistic lighting
• Path Tracing for global illumination

Learning Outcomes: Advanced knowledge in physical-based rendering, shader programming, and optimization techniques for realistic scene depiction. Developed proficiency in implementing complex material models and light interaction algorithms.

Future Works: Exploration into real-time rendering techniques, further optimization of rendering algorithms, and integration of machine learning models for predictive rendering.

Physic's simulator

Project Overview:
Detailed examination of the physics-based simulation techniques applied to various materials and fluids, highlighting key implemented features and the specific challenges encountered.

Implemented Features:

• Cloth Simulation: Elastic forces according to Hooke's law and damping forces.
• Numerical Integrators: Explicit Euler, Midpoint, and Semi-Implicit Euler for time-stepping accuracy.
• Rigid Body Dynamics: Simulations based on Young's modulus, Poisson's ratio, density, and, for contact interactions, stiffness and damping.
• Fluid Simulation: Implementation of "Position Based Fluids" (2013) by Macklin and Müller, incorporating rest density, density correction iterations, artificial viscosity, and artificial pressure strength.

Challenges: The complexities of accurately simulating physical behaviors in a computationally efficient manner posed significant challenges, especially in achieving stability and realism in cloth and fluid simulations.

Future Works: Further optimization of simulation algorithms, exploration into more complex interaction models between different materials, and incorporation of machine learning techniques for predictive simulations.

View Project Code (soon)

PFE - AR navigation

Project Summary: Bachelor's capstone project focused on developing an AR navigation app leveraging Wi-Fi routing for indoor guidance.

Keywords: AI, Wi-Fi Routing, AR, Unity, Photogrammetry

Project Goal: To create an intuitive AR navigation system that uses Wi-Fi signals for indoor pathfinding, enhancing user experience in complex indoor environments.

Main Components:

• AI-driven pathfinding algorithms for efficient routing.
• Integration of AR technology for immersive navigation cues.
• Use of Unity for app development, ensuring a robust and interactive application framework.
• Employment of Photogrammetry for accurate mapping of indoor spaces.

My Contribution: Played a key role in developing the AI algorithms for Wi-Fi based routing, implemented AR visualizations in Unity, and contributed to the photogrammetric mapping of environments.

Future Work: Plans to refine the AI algorithms for even more precise navigation, expand the app's mapping capabilities to cover larger and more complex environments, and enhance the AR experience with more interactive elements.

View Project Code (soon)

Detection d'anomalie de coeurs avec vision par ordinateur

Project Overview: Computer vision project focused on detecting heart anomalies through advanced imaging techniques.

Keywords: Self-Supervised Learning, Adam Optimizer, Data Augmentation, Segmentation

Project Goal: To accurately identify and classify heart anomalies from medical imaging data, improving diagnostic processes and patient outcomes.

Implementation Highlights:

• Utilized self-supervised learning techniques to efficiently train models with limited labeled data.
• Implemented the Adam optimizer for robust and fast convergence in training deep learning models.
• Applied data augmentation strategies to enhance the model's ability to generalize from medical images.
• Developed segmentation algorithms to precisely locate and identify anomalies within the heart.

My Role: Led the development of segmentation algorithms, contributed to data preprocessing and augmentation techniques, and optimized model performance using the Adam optimizer.

Future Directions: Plans to explore more sophisticated self-supervised learning frameworks, extend the project to include a broader range of heart conditions, and integrate the solution into clinical diagnostic workflows for real-time applications.

View Project Code (soon)