Murat Ambarkutuk, Ph.D.
Senior Perception Engineer
Computer Vision • 3D Reconstruction • Robot Perception
Learn More
Senior Perception Engineer
Computer Vision • 3D Reconstruction • Robot Perception
Learn MoreI am a Senior Perception Engineer at RSquared Robotics, a division of Ranpak, where I design, build, and deploy computer vision and machine learning systems for real-time industrial automation. My work spans sensor fusion, 3D reconstruction, deep learning acceleration, and production-grade inference systems.
Virginia Tech
Robot perception, multi-sensor fusion, 3D reconstruction, large-scale aerial mapping
Virginia Tech
Wireless propagation modeling, indoor localization techniques
Kocaeli University
Embedded systems, robotics, machine vision, sensing platforms
My research focuses on large-scale 3D perception systems that integrate geometry, learning, and optimization. I work at the intersection of multi-view computer vision, 3D reconstruction, and high-performance deep learning.
Robust camera pose estimation, bundle adjustment, COLMAP pipelines, and large-scale aerial reconstruction.
TSDF fusion, voxel hashing, multi-view stereo, and neural scene representations including 3D Gaussian Splatting.
TensorRT optimization, CUDA kernels, quantization, operator fusion, and deployment on embedded hardware.
Model distillation (DINOv2, ViT to YOLO), multi-class segmentation, robust tracking under challenging conditions.
Multi-view depth fusion, photometric consistency, winsorized regression, confidence weighting.
High-fidelity Gaussian Splatting, splat compression, differentiable rendering, cross-view feature fusion.
Full perception stack rewrite for high-throughput packaging automation; multi-camera calibration, YOLO-based detection, and TensorRT-accelerated inference.
Multi-class segmentation pipeline for on-line carton graphic validation and hazard-label inspection in industrial settings.
Header-only, cross-platform ONNX/TensorRT engine with dynamic shapes, multi-GPU support, CUDA stream parallelism, and unified logging/configuration.
Winsorized multi-view depth fusion with polynomial regression weighting and robust confidence filtering; production-grade C++/CUDA implementation.
Real-time TSDF integration using voxel hashing; supports depth-only and depth+color fusion with GPU-optimized memory access for large-scale mapping.
Prototypical-network architecture using a frozen DINOv2 backbone and raster logo support sets for robust industrial brand and logo recognition.
Distilling high-capacity DINOv3 features into lightweight YOLOv8 models to achieve real-time, high-accuracy detection and segmentation on embedded hardware.
Evaluation of classroom engagement using advanced localization and sensing techniques.
Radio wave propagation approximation for indoor localization of mobile robots.
Automatic road-profiling system with advanced motion control and embedded sensing.
Remote tracking and monitoring of underground environments via wireless sensor networks.
Mecfuture – Internet-based mechatronics laboratory enabling remote experimentation and teaching.