ImaginSynth: Bridging the Gap Between 3D Workflow and Machine Learning with Open Source Synthetic Data
Tejas Patel, Sandeep Shivam, Amit Kumar Padhy, Bharadwaj Vulugunda, Chaitanya Kulkarni, Chandrashekhar Medicherla
Abstract
This paper explores the realm of synthetic data, emphasizing its crucial role in addressing the extensive data requirements of computer vision in deep learning. Synthetic data, created rather than collected, offers a unique solution to democratize access to diverse datasets, which is essential for robust model training. The paper introduces ImaginSynth, an open-source framework developed in Python, designed for creating synthetic data by leveraging the capabilities of Blender, a popular 3D toolset. Emphasizing accessibility and readability, ImaginSynth acts as a bridge between the mature tools of the 3D workflow and machine learning frameworks. The authors make a compelling case for the significance of open-source synthetic data in mitigating issues such as fairness and bias. Additionally, the paper delves into the impact of domain randomization on synthetic training data, conducting experiments to fine-tune a convolutional neural network (CNN) on small synthetic datasets and assessing its performance on real images. The study sheds light on the potential of synthetic data tools, like ImaginSynth, to revolutionize the intersection of 3D workflows and machine learning, contributing to the advancement of computer vision technologies.