Mehrshad Kafi

• Throughout my work experience and PhD program, I have gained expertise in 3D deep learning and computer vision, with a focus on state-of-the-art models like Mesh R-CNN and SynSin. 
• I have explored various concepts such as Neural Radiance Fields, controllable image synthesis, differentiable volumetric rendering, and object pose detection. Additionally, I have practical experience in handling diverse 3D data formats, implementing camera models, and working with rotation representations. 
• I have also contributed to the field of neural network acceleration and mentored graduate students. Overall, my work showcases a comprehensive understanding of advanced 3D deep learning techniques, practical implementation skills, and leadership abilities.

RESEARCH ENGINEER & RESEARCH ASSISTANT                                                             

McMaster University | Hamilton | 2017 – Present

Object Detection in 3D: Explored Mesh R-CNN, a state-of-the-art model combining image segmentation and 3D structure prediction tasks to extend object detection beyond 2D by generating 3D meshes of detected objects utilizing its elements: Voxels, meshes, graph convolutions, and Cubify operators.

View Synthesis in 3D: Hands-on practice on SynSin, the latest state-of-the-art view synthesis model for 3D deep learning and its applications in AR, VR, gaming, and more.

Human Body: Experiences in formulating its 3D modeling problem, applying and using Linear Blend Skinning technique and the SMPL model, estimating 3D human pose and shape using SMPLify algorithm.

Neural Radiance Fields (NeRF): Explored and trained NeRF as a breakthrough approach to representing 3D scenes, enabling the rendering of a single scene from various angles, capturing fine geometry and texture while requiring less disk space compared to traditional methods.

Controllable Neural Feature Fields: Explored controllable 3D-aware image synthesis using the GIRAFFE model, integrating concepts from NeRF, GANs, and 2D CNNs, and gained proficiency in generating feature fields and mapping them to RGB images for controllable scene generation.

Differentiable volumetric rendering: Experiences in it using voxel 3D data representation, providing advantages over mesh models and using concepts of ray sampling, volume sampling, and ray marching, with practical implementations in PyTorch3D and Demonstrated coding of reconstructing 3D object shapes and colors from multi-view images using differentiable volumetric rendering.

Object pose detection and tracking: Explored and implemented using differentiable rendering, formulating pose estimation and optimizing object pose based on image observations by using Pytorch3D.

Point clouds to mesh models: Implemented a project focused on fitting deformable mesh models to raw point cloud observations and surface reconstruction utilizing PyTorch optimization and various loss functions such as Chamfer distances and mesh edge loss.

Experienced in other 3D computer vision practices: 

• Rendering, rasterization, and shading such as Lambertian and Phong
• Rotation representations and conversions, exponential mapping and log mapping in the SE(3) space.
• 3D data formats, point clouds, meshes, and voxels, and camera models and coordinated systems.

Neural Network Acceleration: Conducted research in neural network (NN) compression techniques especially in mixed precision quantization.

Compression and Encryption: Optimized the compression and encryption process by developing optimization algorithms, resulting in an 8-fold acceleration and investigated concepts and apps of Deep Reinforcement Learning for the PhD comprehensive exam.

TEAM LEAD

Iran Telecommunication Research Center | Tehran | 2015 – 2017

Led a team of engineers in researching IoT technologies while developing a 10-year roadmap for the national deployment of Connected-Cars, and Designed and implemented the backend and frontend of a software solution for communications link budget analysis.