Image Processing for Deep learning Training

About this project

This project involved building a synthetic data generation pipeline using a precompiled Unreal Engine running the ‘Clara’ server via Docker, which provided a realistic 3D traffic simulation environment. I captured both labeled and unlabeled images, annotating objects such as vehicles and pedestrians. The images were then converted to raw or pre-processed formats, optimizing them for AI model training.

I developed a Python Tkinter tool to interact with the simulator, allowing configurable image capture, including sequential frames for video-based datasets.

Additionally, used tensorflow and algorithms or machine learning approaches to convert a real raw image into a processed image, making sure to programatically take care of each pipeline stage (tone mapping, bayer filter demosaicing, etc).

Also created a simple python tkinter tool which interacts with the simulator to capture images with lots of configuration. Including sequential images, which makes sure that it takes multiple shots of the same moment, so that each frame is the image for video format type of training.

Key Features

• Synthetic data with Unreal Engine and Docker
• 3D traffic simulation
• Automated object annotation
• Image conversion for AI training
• Tkinter
• Image processing pipeline