This repository contains the code, models, and results from my engineering thesis titled "Computer Vision Methods for Detection and Tracking of Advertising Logos in Videos". The thesis was submitted at Wroclaw University of Science and Technology (Politechnika Wrocławska). The whole project including research, model training, result analysis, and thesis writing, has been completed within a three-month period (October - December 2024).
- Thesis file - Contains the finished thesis document (available only in Polish).
- Models - Includes all models trained for logo detection.
- Python-app - A Python-based application that allows users to select a video and run logo detection using the trained models.
- Training-results - Contains image files showcasing training results and the performance of models under various disruptions.
The objective of this project was to develop and evaluate computer vision methods for detecting and tracking advertising logos in videos. The approach involved:
- Initial training of YOLOv5, YOLOv8 and RT-DETR models and assessing their aptitude for the task. The models have been trained on the Kaggle platform using Ultralytics tools.
- Preparing custom dataset for logo detection. Initial portions of the dataset were labeled by hand using the Roboflow website. In later steps the labeling process was automized by using the initially trained model.
- Training deep learning model for object detection and logo recognition. RT-DETS was used as base and fine-tuned in several training iterations with different hyperparameters and dataset versions.
- Developing an application to apply trained models on real-world videos.
- Python (OpenCV, Tkinter, Asyncio)
- Ultralytics framework for conducting model training, adjusting hyperparameters and gathering training results
- Image detection models based on CNNs (YOLOv5, YOLOv8) and Transformers (RT-DETR)
- Kaggle notebooks for runing model training on external GPU
- Roboflow website for labeling the data and applying preprocessing and augmentation on the dataset
This project has been prepared solely for academic purposes, and no monetary gains have been made from it. All components of this project are free to use in non-commercial projects.
