AmirHossein (Amir) Razlighi
Computer Vision Researcher | 3D Vision & Graphics | Generative Models
About Me
I'm currently an MSc student in Computer Science at UCY, while researching in 3D Computer Vision and Generative AI at GrUVi lab @ SFU. I am very interested in how computers perceive the world in 3D, and how we can generate realistic worlds using machine learning! I am always open to discussions and connections on research, projects, and potential collaborations. Please feel free to reach out if you want to chat about anything related to computer vision, machine learning, or just want to say hi!
Contact:
Education
Research Experiences
Working Video Generative models and 3D Computer Vision.
Continued the project started at ETH. Worked on Dynamic 3D Reconstruction using Signed-Distance Functions (SDFs) integrated into the Gaussian Splatting pipeline. Created a dataset of challenging deformable scenes (breaking sphere, growing plant, etc.) and designed a complete pipeline for 3D scene encoding and geometry/color reconstruction.
Worked on NeRF models for 3D reconstruction of dynamic scenes. Studied SDFs to disentangle geometry and appearance. Developed morphing between two SDFs based on RGB images and designed SDF-based regularizers for continuous morphing learning in inverse-rendering.
Worked on vision-based robotic manipulation. Studied 3D scene encoding and reconstruction using NeRF models; implemented and customized SRT (Scene Representation Transformer). Customized the ARNOLD dataset for canonical cube views of robot scenes (LINK). Worked on Google's PerAct model and implemented custom agents in IsaacSim.
Industrial Experience
Working on ETA and Navigation for ride-hailing.
Worked on and supervised the full process (from R&D to production) of two projects:
- First member of the R&D team "Tapsi Lab," using computer vision to improve ride-hailing UX.
- Built the Driver-Arrived AR project from scratch, detecting driver position via camera + GPS integration.
- Used TensorFlow.js to run the complete AR experience entirely client-side.
- Created a customized version of AR.js with custom rendering and data-integration layers.
- Collaborated cross-functionally with design, management, business, and DevOps teams.
Publications
Confident Splatting: Confidence-Based Compression of 3D Gaussian Splatting via Learnable Beta Distributions
AmirHossein Naghi Razlighi, Elaheh Badali Golezani, Shohreh Kasaei
3D Gaussian Splatting enables high-quality real-time rendering but often produces millions of splats, resulting in excessive storage and computational overhead. We propose a novel lossy compression method based on learnable confidence scores modeled as Beta distributions. Each splat's confidence is optimized through reconstruction-aware losses, enabling pruning of low-confidence splats while preserving visual fidelity. The proposed approach is architecture-agnostic and can be applied to any Gaussian Splatting variant. In addition, the average confidence values serve as a new metric to assess the quality of the scene. Extensive experiments demonstrate favorable trade-offs between compression and fidelity compared to prior work.
N4DE: Neural 4D Evolution under Large Topological Changes from 2D Images
AmirHossein Naghi Razlighi, Tiago Novello, Asen Nachkov, Thomas Probst, Danda Paudel
In this work, we address the challenges in extending 3D neural evolution to 4D under large topological changes by proposing two novel modifications. More precisely, we introduce (i) a new architecture to discretize and encode the deformation and learn the SDF and (ii) a technique to impose the temporal consistency. (iii) Also, we propose a rendering scheme for color prediction based on Gaussian splatting. Furthermore, to facilitate learning directly from 2D images, we propose a learning framework that can disentangle the geometry and appearance from RGB images. This method of disentanglement, while also useful for the 4D evolution problem that we are concentrating on, is also novel and valid for static scenes.
MARS: Multi-task Action Prediction for Robot Manipulation based on Scene Representation
Haoping Xu, Richard Hanxu*, AmirHossein Naghi Razlighi*, Alan Aspuru-Guzik, Florian Shkurti, Animesh Garg
This paper presents MARS (Multi-task Action Prediction for Robot Manipulation based on Scene Representation), a novel approach to robot manipulation tasks using a transformer-based framework. MARS leverages Scene Representation Transformers (SRT) to render orthogonal novel views from RGB multi-view observations, enabling the prediction of the next key frame's best end-effector pose and gripper state. The model is trained using the PerAct synthetic dataset, which includes expert demonstrations for a variety of tasks. By conditioning on language descriptions and utilizing keyframe-based behavioral cloning, MARS effectively generalizes across multiple tasks, demonstrating robust performance in complex robotic manipulation scenarios.
Projects (GitHub)
Tiny NeRF
A PyTorch implementation of tiny NeRF (Neural Radiance Field). A much smaller version of the original that can be trained even on CPU in a reasonable time.
RayTracing from Scratch
Basic ray tracing concepts implemented in Python using only NumPy and Matplotlib, for understanding and experimenting with fundamental ideas.
Marching Squares
Python implementation of the Marching Squares algorithm for 2D terrain generation, with binary/float value systems and various noise generators.
Semantic Scholar Search Engine
Three-phase search engine: (1) TF-IDF index + compression, (2) ML-based ranking (Naive Bayes, Neural Nets, LMs), (3) web scraping + Streamlit frontend.
BlurSim
Introduces a new loss function for accurately comparing differences between two blurred images.
CMinus Compiler
CMinus-to-Python compiler with four phases: scanner, parser, error analysis, and semantic analysis.
Teaching & Volunteering
Teaching Assistant — Sharif University of Technology
Oct 2021 – July 2025 | All positions are voluntary / unpaid.
- Adv. 3D Computer Vision (Graduate) — Quizzes on multi-view reconstruction; lecture series on NeRF and Gaussian Splatting. Prof. Shohreh Kasaei
- Modern Information Retrieval (Head of Project) — Led 3-phase final project: Classical IR, ML in IR, Deep methods + LLMs. Dr. Mahdieh Soleymani
- Machine Learning (Head TA, Project Section) — Managed team of 10; CV and NLP final project. Dr. Fatemeh Seyed Salehi
- Convex Optimization — Convex Functions questions; managed Duality homework grading. Dr. Amir Najafi
- Digital Image Processing (Graduate) — Deep learning slides and teaching sessions; quizzes on image transformations and morphological operations. Prof. Shohreh Kasaei
- Scientific and Technical Presentation (Head TA) — Prof. Shohreh Kasaei
- Artificial Intelligence — Computer Vision project designer. Dr. M. H. Rohban
- Artificial Intelligence — Head of final project design. Dr. M. H. Rohban, Dr. M. Soleymani
- Fundamentals of Computer Vision (Head of Workshops) — Exercises on 3D Vision. Prof. Shohreh Kasaei
- Database Design — Dr. M. Varmazyar
- Linear Algebra (2 semesters) — Re-designed all slides; theoretical exercise design. Dr. Maryam Ramezani
- Probability and Statistics — Dr. Mahdi Jafari
- Advanced Programming in Java (3 semesters) — Held 3 OOP workshops; designed GUI phase of course project.
Other Volunteering Activities
- Vice President of Students' Scientific Chapter (SSC)
- Head of Technical Beta Team (8th WSS) — Created an R&D chatbot trained on FAQ data to help conference participants get instant answers. GitHub
- Founder of Byte Publication — LinkedIn
- Enrolled in "RTI" Program (Oxford University)
- Member of "3D Vision" Reading Group (University of Toronto)
Hobbies
Outside of work and studies, I enjoy Playing Video Games and Reading Science Fiction Novels! You can find me via below links: