A Master's project for training, comparing, and deploying state-of-the-art deep learning models for object detection in satellite imagery. This framework includes a complete pipeline from data preprocessing to model evaluation and final deployment in a QGIS plugin.
The primary goal of this research is to evaluate and compare the performance of several modern object detection architectures on the complex task of identifying objects in satellite imagery. The project handles the entire machine learning lifecycle:
- Data Unification: Combining multiple public datasets (DIOR, DOTA, FAIR1M) into a single, cohesive dataset.
- Data Preprocessing: Handling massive satellite images through a robust tiling mechanism with overlap.
- Model Training: Implementing training scripts for multiple SOTA models that support Horizontal Bounding Boxes (HBB).
- Comparative Analysis: Evaluating models based on standard metrics like mAP to determine their strengths and weaknesses for this specific domain.
- Deployment: Integrating the best-performing models into a custom QGIS plugin for practical use by GIS analysts.
The dataset used in this project is a custom compilation of several public benchmarks:
- DIOR
- DOTA-v2.0
- FAIR1M
The raw data was processed through a pipeline involving class unification, stratified splitting (to ensure class balance), and slicing of large images into 800x800 tiles with overlap to create a model-ready dataset.
This section will be updated with the final results of the model comparison. Key metrics include mAP50, mAP50-95, inference speed, and parameter count.