Esat Kalfaoglu
Verified Expert in Engineering
Computer Vision Developer
Ankara, Turkey
Toptal member since September 1, 2022
Esat obtained his master's degree and is currently finalizing his PhD studies on deep learning topics. He has a strong background in object detection, action recognition, panoptic segmentation, and deep reinforcement learning. Esat is greatly interested in transformer-based architectures.
Portfolio
Experience
- Artificial Intelligence (AI) - 8 years
- Computer Vision - 8 years
- Machine Learning - 8 years
- PyTorch - 7 years
- Deep Learning - 7 years
- Multi-GPU Training - 5 years
- Open Neural Network Exchange (ONNX) - 5 years
- Road Topology Understanding - 3 years
Preferred Environment
PyTorch, OpenCV, Python 3, Linux, NVIDIA TensorRT, Git, WandB, Open Neural Network Exchange (ONNX), Snapdragon Neural Processing Engine (SNPE)
The most amazing...
...thing I've accomplished is the third-ranking position in my department during my bachelor's degree graduation.
Work Experience
Machine Learning Engineer
Togg
- Implemented deep learning-based lane detection algorithms. Also dealt with the deployment of these algorithms to Nvidia Orin and Texas TDA4VH cards.
- Studied conversion of deep learning algorithms to Onnx, TensorRT, TIDL, and SNPE formats.
- Implemented semantic, instance, and panoptic segmentation algorithms for a general urban understanding of autonomous driving.
- Created a development environment with the integration of PyCharm and VS Code into the singularity environment.
- Implemented transformer-based algorithms, where I became familiar with multi-scale deformable attention, a famous architecture in recent studies.
- Shared my experience in end-to-end driving, including new concepts such as the BEV paradigm, prediction, temporal utilization, and sensor fusion.
Team Lead
AutoDidactic Technologies
- Completed the first phase of the HAVELSAN FIVE-ML project tests.
- Designed the OpenAI Gym environment in the HAVELSAN FIVE software to develop reinforcement learning algorithms in it.
- Utilized PPO algorithm for air-to-air and air-to-ground military combat scenarios in FIVE software.
- Improved my project and Git skills as a team lead.
Research Scientist
METU Image Analysis Center
- Passed the three report periods of the project funded by the Scientific and Technological Research Institution of Turkey and the British Council and printed academic studies related to the project.
- Implemented all five computer vision tasks, including indoor person counting, indoor activity estimation, light state detection, window state detection, and curtain openness ratio prediction.
- Obtained and verified an under 10% error for an indoor person counting algorithm in a large environment, where numbers can increase to 150 people for a whole day.
Experience
Learning Virtual Forces with Artificial Intelligence | FIVE-ML
Behavior models managing FIVE entities depend highly on the rule-based behavior developed by field experts and system engineers. However, the rule-based operation of FIVE software requires intensive programming and field experts' guidance. Hence, the process is highly labor intensive. Furthermore, the complexity and burden of this task increase significantly with the complexity of the scenario. In addition, virtual entities with rule-based behavior have standard and predictable reactions to their environments. Therefore, in this study, we present the transition studies from rule-based behavior to learning-based adaptive behavior via reinforcement learning techniques coupled with other machine learning techniques, namely the FIVE-ML project. For this aim, reinforcement learning-based behavior models are trained for air-to-air and air-to-ground scenarios with up to six virtual entities. It is observed that virtual entities trained with reinforcement learning dominate existing rule-based behavior models.
Intelligent System for Building Energy Retrofitting | SISER
OBJECTIVES
• Create a 3D spatial thermal model of existing buildings.
• Collect temporal information on occupants, such as their numbers and energy activities, and their energy use behavior, including lighting changes, open and close detection of curtains, shades, and doors. All information is obtained from IP cameras in an automated way.
• Develop methods to capture stakeholder information and occupant comfort information.
• Create methods to streamline and automate retrofit scenario development and simulation-based analysis.
• Develop the collaborative decision-making framework to address energy performance and comfort to the satisfaction of all stakeholders.
• Test and validate SISER through a case study.
Education
Ph.D. Degree in Multimedia Informatics
Middle East Technical University - Ankara
Master's Degree in Electrical and Electronics Engineering
Middle East Technical University - Ankara, Turkey
Bachelor's Degree in Electrical and Electronics Engineering
Bosphorus University - Istanbul, Turkey
Student Exchange Program in Electrical and Computer Engineering
The University of Texas at Austin - Austin, TX, USA
Skills
Libraries/APIs
PyTorch, OpenCV, NumPy, Gradio, PyTorch Lightning
Tools
Git, Open Neural Network Exchange (ONNX), Snapdragon Neural Processing Engine (SNPE), MATLAB, OpenAI Gym, Tmux, PyCharm
Languages
Python 3, Java, Python
Frameworks
Flask
Paradigms
High-performance Computing (HPC)
Platforms
Linux, MacOS, NVIDIA Jetson AGX Orin
Storage
Amazon S3 (AWS S3)
Other
Deep Learning, Computer Vision, NVIDIA TensorRT, WandB, Image Processing, Digital Signal Processing, Linear Algebra, Probability Theory, Algorithms, Deep Reinforcement Learning, PyCUDA, Machine Learning, Artificial Intelligence (AI), Natural Language Processing (NLP), Gradio.app, Autonomous Navigation, Advanced Driver-assistance Systems (ADAS), Generative Adversarial Networks (GANs), Generative Pre-trained Transformers (GPT), Research, Self-driving Cars, ONNX Runtime, TIDL, Road Topology Understanding, Centerline Detection, Multi-GPU Training, Multi-Node Training, Transformer Models, CARLA, Sensor Fusion, Autononous Driving, 2D Object Detection, 3D Object Detection, Lane Detection, Traffic Element Detection
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
Share your needs
Choose your talent
Start your risk-free talent trial
Top talent is in high demand.
Start hiring