Scene analysis and data augmentation

In the last few decades, image sensors attached to airborne platforms, UAVs or satellites have evolved making possible the capture of high-resolution multi-spectral images. As a result, remote sensing object recognition and classification has become a widely-studied field providing solutions on the detection and classification of buildings, other structures, land and air vehicles. As a consequence, many remote-sensing applications, such as city planning, urban mapping and urban change detection have progressed by using object detection systems that employ aerial images and reconstructed 3D representations. As there is a lack of annotated real data, an important aspect of the project was the creation of a data generation system supporting fully 3D scenes and data models in order to produce rendered annotated images and different background environments.

This project is motivated by generating training data in order to obtain high performance object detection and recognition. It is also focused on the identification of the best existing methods for scene analysis and data augmentation. It aims to produce more realistic 3D training data to ensure the delivery of systems addressing specific scene analysis requirements. In light of those results, this project will offer solutions to improve the quality of the generated data and the overall performance of the object detection and recognition solutions by enhancing data through a pre-processing stage, supporting multi-scale object recognition, and offering a data generation tool supporting full multidimensional rendering.

The general aim is to offer further functionalities and features allowing data generation for more advanced scenarios covering all the scenarios and applications that are related to scene analysis.