Sam (Seyed M.) Hosseini

Welcome!

I am a Computer Vision Researcher & Engineer at Elder Lab, York University, specializing in 3D scene understanding and metric depth estimation under the supervision of Prof. James Elder. My work bridges the gap between mathematical theory and production, integrating geometric reasoning with robust deep learning pipelines.

Prior to this, I completed my B.Sc. in Electrical Engineering at the University of Tehran (2021), where I conducted research on single-view 3D reconstruction of symmetrical objects at the Computational Audio-Vision Lab with Prof. Reshad Hosseini.

Core Interests

3D Reconstruction & Multi-View Geometry
Geometric Deep Learning
MLOps & Scalable Cloud Deployment
Medical Image Analysis

View my CV here. Have questions about my background or research? Chat with my AI Assistant

Current Research

My research explores the frontier of 3D scene understanding, blending classical geometric constraints with modern generative approaches. I focus on developing Semantics-Guided Zero-Shot Perception systems that leverage vision foundation models to recover scale-accurate metric depth from single images. I am also investigating the synergy between Gaussian Splatting and Diffusion Models for high-fidelity single-view 3D reconstruction, aiming to resolve scale ambiguity and improve geometric consistency from limited observations.

Selected Projects

A showcase of my engineering and applied research work.

MonoDiffSplat: 3D Scene Reconstruction using 2D Gaussian Splatting and Video Diffusion Inpainting

Combined Depth Anything V3 and G4Splat pipelines and introduced a parallax-inducing camera trajectory strategy for the generative inpainting stage. These trajectories force the video diffusion model to hallucinate valid depth cues and resolve disocclusions by actively moving into the scene

VisualRAG: Agentic Competitive Intelligence

Engineered an autonomous Agentic RAG system using DeepSeek-OCR and LangGraph. Treats research as a visual task to extract structured data from complex PDFs, solving retrieval bottlenecks where standard text scrapers fail.

Semantics-Guided Zero-shot Metric Depth

Integrated vision foundation models with state-of-the-art estimators to achieve scale-accurate quantitative outputs for metrology. Designed semantic-guided geometric scaling for unconstrained environments.

Real-Time Detection & MLOps Pipeline

Engineered a complete pipeline for football player detection. Fine-tuned YOLOv11, containerized with Docker, automated CI/CD via GitHub Actions, and deployed on AWS.

Single-View Geometric Reconstruction

Formulated a geometric reconstruction algorithm for symmetric objects from single images. Benchmarked against NeRF and Gaussian Splatting methods (One-2-3-45, DreamGaussian) on Pix3D.

3D Fiber Reconstruction from Medical Data

Built a CNN–Transformer framework for 3D orientation estimation and deterministic fiber tracking, applying transferable geometric modeling techniques for robust quantitative analysis.