Automated Dataset Generation with Blender for Deep Learning-based Object Segmentation

Abstract

Listen

Translate article

"Give a man a synthetic dataset and you help him for a project; teach a man to generate his own dataset and you help him for a lifetime." Developments in digital technology led to an unprecedented increase in the amount of data in industry and in science as well, including the field of computer vision. This enabled the evolution of convolutional neural networks (CNNs) and other deep learning (DL) architectures that require a large amount of data for their training. Even though the large-scale collection of visual data is relatively cheap and easy to implement, the annotation of the collected data is a time-consuming process. Although, with the use of pre-trained models only a small-scale fine-tuning dataset is required, the labeling of only a few hundred or few thousand images for object segmentation or instance segmentation is still a laborious task, if done manually. Recent trends suggest that synthetic data will play a significant role in the future of DL and there are numerous research papers promoting various synthetic datasets and/or models trained on them. Unfortunately, there is much less information on how to create such a dataset and what are the important points one has to keep in mind when creating their own. In this paper, we summarize some of the available tools for creating a synthetic dataset for object segmentation and provide a detailed example using the Blender 3D creation suite. During our example, we attract attention to the vital points and considerations for automatic dataset generation for object segmentation.