A 3D Synthetic Image Generation Framework: Enabling Deep Learning Feature Detection and Object Classification

Data scarcity represents a significant challenge in deep learning, particularly in novel applications where acquiring and annotating large datasets is costly and time-consuming. This thesis explores the use of synthetic data generation as a means to mitigate these limitations, focusing on its application to loss prevention in retail environments. The work is structured into two key contributions, the development of a python library named Synthetic Generation Framework, and the design of an object recognition model based on local features. The Synthetic Generation Framework, built using the Blender Python module, was designed to create realistic 3D datasets equipped with ground truth annotations, in order to accelerate model development and testing. This tool enabled efficient prototyping of an object recognition system that detects grocery products in self-checkout scenarios by leveraging local feature descriptors; the system is built upon KNIFT, a deep-learning-based local feature descriptor designed for robust keypoint matching. This work demonstrate the potential of synthetic data generation not only as a means to address data scarcity, but also as a useful tool for rapidly developing and validating AI-driven applications.