Desenvolvimento de extensão para geração de conjuntos de dados sintéticos

Abstract

One of the main challenges in developing artificial intelligence models based on deep neural networks is obtaining high-quality data for training datasets. As tasks become more complex, the costs of annotation and the difficulty of acquiring data increase, making the process more time-consuming and expensive. As an alternative, researchers have explored the use of virtual environments to automate the annotation of low-cost, error-free datasets. In this context, this work presents the development of an extensible Python plugin for the open-source software Blender, which uses the graphics engine to generate photorealistic synthetic datasets aimed at computer vision challenges such as segmentation and bounding box detection. In the end, the tool was used to create a dataset with retail product images, with the goal of evaluating the quality of the generated annotations and the impact of synthetic data on training. The Faster R-CNN model with ResNet50 backbone, pre-trained on the COCO dataset, was trained on a set of 10,000 images containing 5 product classes, achieving a mAP@0.5:0.95 of 48,1% on the synthetic validation set and a mAP@0.5:0.95 of 24,8% on the real validation set.