I build computer vision systems that actually ship — from dataset creation and annotation tooling all the way to TensorRT-optimized inference running on factory floors and vehicles.
name: Mehmet OKUYAR
role: AI & Software Development Manager @ Pro Sicht
education: BSc ×2 (Sakarya Univ. of Applied Sciences · Óbuda University, Budapest)
MSc, Computer Engineering
based_in: Sakarya, Türkiye
focus: [ Computer Vision, Deep Learning, Edge AI, Multimodal Models ]
domains: [ Industrial Inspection, Automotive & ADAS, Medical Imaging, Smart City ]
ask_me_about:
- Object detection & tracking (YOLOv4 → YOLOv11, DeepSORT, ByteTrack)
- Semantic / instance segmentation (U-Net, SAM, YOLO-seg)
- Model deployment & acceleration (TensorRT, ONNX, CUDA, Jetson)
- Data-centric AI: auto-labeling, augmentation, dataset engineering
currently: Vision-based quality inspection & multimodal LLM pipelines
fun_fact: 🐎 Horse riding & traditional Turkish archery|
Automated defect inspection and quality control for automotive manufacturing. Industrial cameras, real-time TensorRT inference, PLC/line integration and PyQt operator interfaces built to survive a production shift. |
Lane segmentation, traffic sign detection, vehicle counting, speed estimation and zone analytics — plus depth-aware detection with ZED stereo cameras for real-world distance measurement. |
|
Open-source utilities that remove the grind from dataset prep: auto-labeling from pretrained weights, video-to-frame slicing, label-preserving augmentation, tiling for small-object detection, and converters between YOLO / COCO / mask formats. |
Vision-language models (Qwen-VL), generative approaches to data synthesis, 3D reconstruction and Gaussian splatting, and peer-reviewed work on medical image segmentation and classification. |
Languages
Deep Learning & Vision
Deployment & Acceleration
Backend & Apps
Vision Hardware
Star and fork counts update automatically.
|
Fuses YOLO detections with ZED stereo depth so every detected object comes back with a real-world distance — the building block for obstacle avoidance and robotics perception. |
|
|
YOLOv7 + DeepSORT pipeline that counts vehicles, holds identity across frames, estimates speed and reports per-zone traffic statistics. |
|
|
Annotation is the bottleneck in every CV project. This tracks your chosen object through a video and writes the images plus YOLO |
|
|
Traffic sign detector trained on a custom dataset and exported to TensorRT for real-time inference on embedded GPUs. Pairs with the TTVS Turkish Traffic Sign Dataset. |
|
|
Draw counting regions interactively, drag them around live, and get per-region object counts — no config files, no restarts. |
|
|
Lane detection model trained on a self-collected dataset — the perception layer for lane-keeping and drivable-area estimation. |
|
| Tool | What it solves |
|---|---|
| Auto_video_label | Tracks an object through a video and auto-writes YOLO .txt labels — dataset creation without the click fatigue |
| Slicing-Image-With-Label | Tiles large images with their labels so small objects survive the resize to model input size |
| Augmentation_for_Yolo_labeling | Augments images and their YOLO-format annotations together, keeping boxes valid |
| Convert-Mask2Yolo | Converts segmentation masks into YOLO-seg polygons — reuse old datasets on new architectures |
| yolobbox2polygon | Promotes YOLO bounding boxes to segmentation polygons using SAM-HQ |
| VideoSlicing | Turns raw footage into frame datasets with controllable sampling |
| Train_Test_Split_txt | One-shot train/test split in YOLO's expected .txt manifest format |
| convert_dcim2png | Converts DICOM medical images to PNG for standard vision pipelines |
| ubuntu-libs-kurulum | Scripted CUDA/cuDNN/DL stack setup on Ubuntu — bare metal to training, fast |
NOW ● AI & Software Development Manager — Pro Sicht
│ AI-powered visual inspection and measurement systems for production
│ lines: full vision pipeline from camera to operator screen.
│
● MSc, Computer Engineering — Sakarya University of Applied Sciences
│ Deep learning for medical image segmentation & classification.
│
● AISET Autonomous Vehicle Community
│ Perception stack for autonomous driving: lane segmentation, traffic
│ sign detection, embedded TensorRT deployment.
│
● Yapay Zeka Araştırma Merkezi (YAZEM) — AI Research Center
│ Applied computer vision research.
│
● Toyota Motor Manufacturing Turkey
│ Automotive manufacturing & quality processes.
│
● Ford Otosan — Intern
│ Freespace (drivable-area) semantic segmentation for highway imagery.
│
● BSc, Electrical & Electronics Engineering — Óbuda University, Budapest
│
● BSc, Electrical & Electronics Engineering — Sakarya Univ. of Applied Sciences
Segmentation and classification of skin burn images with artificial intelligence: Development of a mobile application Burns (Elsevier), 2024 — Yıldız M., Sarpdağı Y., Okuyar M., et al. Segmentation, classification and object detection on skin burn images, delivered as a mobile application for burn-degree assessment. → ScienceDirect
Ischemia and Hemorrhage detection in CT images with hyperparameter optimization of classification models and improved U-Net segmentation model Sakarya University Journal of Computer and Information Sciences (SAUCIS), 2023 — Okuyar M., Kamanlı A. F. Hyperparameter-optimized classifiers combined with an improved U-Net for stroke-related lesion segmentation in CT. → SAUCIS
I write about applied computer vision on Medium:
- Plastik Yüzeylerdeki Kusurların Tespiti: Sanayideki Önemi ve Yapay Zeka Çözümleri
- Leveraging Text-to-Image Generation Models for Enhanced AI Data Collection
- How Artificial Intelligence is Secretly Present in Our Daily Lives
I'm open to computer vision research, industrial AI deployments, and open-source CV tooling. If one of the tools above saved you time, a ⭐ helps others find it too.
"Data is the largest factor in deep learning — the better the dataset, the better the model generalizes."