Arda Öztüner

About Me

I am a machine learning and deep learning enthusiast with a keen interest in computer vision and image processing. My expertise lies in developing and deploying neural networks to solve challenging problems, from medical image analysis to object detection.

I have hands-on experience with designing various neural networks architectures, implementing cutting-edge models, and optimizing performance for real-world applications. I am driven by curiosity and enjoy exploring new techniques to push the boundaries of what technology can achieve.

Technical Skills

Python

PyTorch

TensorFlow

Computer Vision

Image Processing

Object Detection

Segmentation

Model Deployment

Transfer Learning

Data Processing

Pandas

NumPy

Scikit-learn

SQL & Database Managment

Education

SEPTEMBER 2023

Artificial Intelligence and Machine Learning Minor Program

Eskisehir Technical University

GPA: 4.0

SEPTEMBER 2022

Computer Engineering

Eskisehir Technical University

GPA: 3.84

My Projects

Computer Vision Object Detection Instance Segmentation Flask App

Fisheye Cameras Road Object Detection

A deep learning-based computer vision project that tackles the challenges of object detection in fisheye camera images. Using the FishEye8K dataset, multiple state-of-the-art models — including YOLOv8, Faster R-CNN, RetinaNet, and YOLOv8-Seg — were evaluated for detecting cars, pedestrians, trucks, buses, and bikes. The project also features a Flask web app for real-time detection on images, videos, and batches with segmentation overlay support.

GitHub

Vision Transformer PyTorch from Scratch Positional Embedding

Vision Transformer (ViT) Implementation from Scratch

This project presents a fully from-scratch PyTorch implementation of the Vision Transformer (ViT) architecture, applied to garbage classification. Inspired by the landmark paper “An Image is Worth 16x16 Words”, the model splits images into patches and treats them as sequences for transformer-based processing. It includes different training strategies (standard, gradient clipping, scheduler) and optimizer comparisons (Adam, SGD, RMSprop), providing a deep insight into ViT training dynamics.

GitHub

CNN Neural Networks Medical Imaging

Brain Tumor Classification

A medical image classification project. This project focuses on classifying X-ray brain scans into Healthy and Tumor categories using both custom CNN architectures and pre-trained models like ResNet50, MobileNetV2, and DenseNet121. Emphasis was placed on preprocessing techniques such as histogram equalization and data augmentation to improve model performance. The best-performing model (Basic CNN 1) achieved 95% accuracy, highlighting the potential of lightweight CNNs in medical applications.

GitHub

Self-Organizing Map Unsupervised Learning Fraud Detection Scratch Implementation

SOM for Credit Card Fraud Detection

This project presents a complete Self-Organizing Map (SOM) implementation from scratch in Python, aimed at detecting credit card fraud using unsupervised learning. It demonstrates how SOMs can cluster application data and flag anomalous behavior without labeled examples. The notebook visualizes fraud-prone regions using a winner map, and the entire implementation is backed by a custom-built testing suite to validate core SOM functionalities.

GitHub

Optical Flow Motion Estimation OpenCV

Optical Flow Visualization with Lucas-Kanade Method

This project demonstrates real-time optical flow estimation using the Lucas-Kanade method. It visualizes object motion between frames using color-coded motion vectors, where hue represents direction and brightness represents speed. Supports both webcam streams and image pairs. Includes customizable parameters for smoothing, stride, and resolution — ideal for understanding dynamic motion in scenes.

GitHub

Data Cleaning Fuzzy Matching Data Visualization

International Seminar Participation Analysis

Cleaned and analyzed a dataset of 1800+ ISUF conference participants. Applied data wrangling techniques including missing value imputation, categorical normalization, and fuzzy string matching (RapidFuzz) for duplicate detection. Performed exploratory data analysis (EDA) and created visualizations to reveal trends by gender, country, title, and year of participation.

GitHub

Blog & Shares

May 18, 2025

Gradient Clipping: Stop the Chaotic Weight Updates!

What is exploding gradients, why does it happen, and how does gradient clipping help? A vital tip for stability in deep learning training.

Learning Rate Scheduling in Deep Learning

May 25, 2025

Learning Rate Schedulers: Smarter Training Over Time

Explore Step Decay, Exponential Decay, and Cosine Annealing methods to optimize how your models learn as training progresses.

June 1, 2025

What Really Measures Success in Object Detection?

mAP, IoU, Precision, Recall — explore the key metrics that define how effective your object detection model truly is.

June 8, 2025

CutMix & MixUp: Smart Tricks for Stronger Models

Explore how mixing images and labels using CutMix and MixUp can boost generalization, fight overfitting, and improve model performance.

May 18, 2025

Understanding SAM: The Segment Anything Model

Discover how Meta AI's SAM is revolutionizing image segmentation. This foundation model enables segmentation using points, boxes, text, and more — even on unseen images. Learn where SAM excels and how it can transform object detection into pixel-level precision.

About Me

Technical Skills

Education

My Projects

Fisheye Cameras Road Object Detection

Vision Transformer (ViT) Implementation from Scratch

Brain Tumor Classification

SOM for Credit Card Fraud Detection

Optical Flow Visualization with Lucas-Kanade Method

International Seminar Participation Analysis

Blog & Shares

Gradient Clipping: Stop the Chaotic Weight Updates!

Learning Rate Schedulers: Smarter Training Over Time

What Really Measures Success in Object Detection?

CutMix & MixUp: Smart Tricks for Stronger Models

Understanding SAM: The Segment Anything Model

Contact Me

Get in Touch