NeurIPS 2026
SemGeo-Gen: Unsupervised Generation of Approximate Cross-Instance Semantic-Geometric Correspondences
PhD Research
Paper and code coming soon
Roy Amoyal
PhD Candidate in Computer Science · 3D Computer Vision
Ben-Gurion University of the Negev
I'm a PhD candidate at Ben-Gurion University under the supervision of Prof. Oren Freifeld, and an Applied Scientist (Research) intern at Amazon.
My research is in 3D computer vision: structure from motion, multi-view stereo, 3D reconstruction, Gaussian Splatting, and 2D/3D registration. Before my PhD, I spent four years building vision systems for autonomous drones.
In my first PhD year (2026), I've been exploring semantic-geometric representations to tackle a hard problem: matching and aligning different objects of the same category. This line of work has already led to two papers at CVPR 2026 and NeurIPS 2026, with a third on the way.
3D ReconstructionGaussian Splatting3D Registration Multi-View GeometryVisual SLAM
News
- Three papers accepted to NeurIPS 2026, including MACRO as a Spotlight.
- Joined Amazon Prime Video Sports as an Applied Scientist (Research) intern.
- Cross-Instance Gaussian Splatting Registration (GSA) accepted to CVPR 2026.
- Gaussian Splashing published in Scientific Reports.
- Wavelet Convolutions for Large Receptive Fields (WTConv) published at ECCV 2024. It has since passed 1,000 citations.
Publications
NeurIPS 2026Spotlight · top 1.3%Amazon Research Internship
MACRO: Training-free Multi-plane Attention for Closeup Render Optimization
NeurIPS 2026
Video Stitching from Multiple Moving Cameras
Paper and code coming soon
CVPR 2026
Cross-Instance Gaussian Splatting Registration via Geometry-Aware Feature-Guided Alignment
PhD Research
GSA aligns two independent 3D Gaussian Splatting models with a similarity transform, even when they depict different objects of the same category, using viewpoint-guided spherical map features and a coarse-to-fine registration.
Scientific Reports 2026
Gaussian Splashing: Direct Volumetric Rendering Underwater
Extends 3D Gaussian Splatting to underwater scenes by modeling scattering and absorption in the medium, for fast, high-quality reconstruction and rendering.
ECCV 20241,000+ citationsUndergraduate Research
Preprints
arXiv 2026
Fast and Memory-Efficient Wavelet Convolutions via I/O-Aware Reformulation
Featured Projects
🏁 Israel's first student autonomous race car: first successful tests · Feb 2025
BGRacing Autonomous: Israel's First Student-Built Autonomous Race Car
I led the autonomous division of BGRacing, Ben-Gurion University's Formula Student team. In 2025 we became the first university team in Israel to build a student-led autonomous race car that drove on its own in real-world tests.
- Led 20 students in 4 teams across the full autonomy stack: simulator, perception, path planning, and control.
- LiDAR-only pipeline (detection → planning → control) that can drive even without state estimation.
- Cone position and distance from a single RGB camera, using known intrinsics and cone size, as a LiDAR backup.
- Extended Kalman Filter fusing GPS, IMU, and wheel speed for state estimation.
Path planning, control & simulation. Built the team's simulator and a full autonomy loop in it: LiDAR cone detection, path planning in the car's frame, and vehicle control. Post
Camera perception. A fast cone detector with distance estimates from a single RGB camera, using known cone size and camera intrinsics. Post
LiDAR perception. Cones detected in Innoviz LiDAR point clouds after removing the ground with robust RANSAC plane fitting. Post
⏩ 1.5× speed · one unedited round
You Only Shoot Once: A Real-Time YOLOv5 Aimbot for Counter-Strike
Built from scratch in 2–3 days while home sick with COVID, before the age of AI coding assistants: no ChatGPT, no Claude.
A real-time computer vision system that plays Counter-Strike: it captures the screen, detects opponents with a YOLOv5 model I trained, and aims and fires automatically. I wrote all of it myself, without a tutorial.
- Trained YOLOv5, one of the newest and fastest real-time detectors at the time, on gameplay frames I labeled, targeting opponents' head and upper body.
- Aimed and fired at detected targets in real time on a single GTX 980 GPU (4 GB VRAM).
- In the round shown, I only moved and listened for enemies. Every kill was made automatically by the system, including every headshot, and it fired about 90% of all shots.
Built as a vision experiment. I don't support cheating in online games.
Experience
-
Applied Scientist (Research) Intern
Amazon · Prime Video Sports
-
Computer Vision Algorithm Engineer
CopterPix Pro · Autonomous Drones
Field testing -
Computer Vision Algorithm Developer Intern
Penta Drone Robotics · Starship program
Certificate (Hebrew)
Honors & Activities
Third place, HackIDC 2021, then Israel's largest student hackathon: 300 students, including MSc and PhD students, from Israel's top universities (Technion, Tel Aviv, Hebrew, and Ben-Gurion).
As second-year undergraduates new to computer vision, my teammates built the app and I designed and implemented its computer vision algorithm.
In the finals, Mark and I pitched to 200+ people and a panel of judges, including:
- Dov Moran, co-inventor of the USB flash drive
- Dr. Dorit Dor, former CTO of Check Point
- Moran Nir, AWS startups lead
Third-place prize - Acknowledged contributor to Understanding Deep Learning (MIT Press, 2023) by Simon J.D. Prince, author of Computer Vision: Models, Learning, and Inference (Cambridge University Press, 2012). Credited in the acknowledgments, page 13.
MIT Press, 2023
Service & Community
- Reviewer for ICCV 2025, CVPR 2026, and NeurIPS 2026.
-
Stack Overflow contributor · 2021–2022Back in the day, before AI assistants like ChatGPT, Claude, and Gemini, developers turned to Stack Overflow for help.
Early in my computer vision journey, I answered questions in its computer vision community on OpenCV, object detection, and image processing, mostly with classical computer vision.
Selected accepted answers
- Real-time detection of a COVID-19 rapid test
- Counting faces in real time on a Raspberry Pi
- Detecting large objects among dense, similar-colored small ones
- YOLOv3 returning zero confidence for every class
- ArUco marker detection returns no IDs
Detecting a COVID-19 test
by color
Tracking faces across frames
on a very low-compute edge device
Separating large objects from dense
small ones: boxes and contours
Teaching
Teaching Assistant, Department of Computer Science, Ben-Gurion University
- Introduction to Deep Learning: neural networks, CNNs, Transformers, PyTorch, and CUDA.
- Introduction to Computer Vision: classical computer vision and modern deep learning approaches.
Education
- PhD, Computer Science, Ben-Gurion University. 3D computer vision.
- MSc, Computer Science, Ben-Gurion University. Autonomous systems and AI.
- BSc, Computer Science, Ben-Gurion University. Data science.
-
First-year Computer Science courses, The Open University of Israel. Independent study during military service.