From Sources to Solutions: Enhancing Object Detection Models through Synthetic Data
Eduard Bartolovic, Tobias Höfer, Clemens Hage, Alfred Nischwitz · Computer Science Research Notes · 2024
Object detection, a fundamental task in computer vision, plays a crucial role in various applications such as autonomous driving, surveillance, and robotics. However, training models for this task require vast amounts of high-quality data, often involving labor-intensive manual labeling. Synthetic data, a promising alternative, remains an active area of research. This paper presents a comprehensive exploration of different object sources for the use of synthetic data in enhancing object detection models. We investigate various synthetic data generation techniques to implant objects into a scene, with a focus on enhancing training data diversity. These objects are either gathered from the training dataset itself using SegmentAnything [1] as a new supervised self augmentation technique or imported from external sources, including a photobox with a rotating table and web scraping of online shops. Moreover, our study delves into the development of a placement logic that gradually evolves from placing objects randomly to placing objects in physically correct orientations to mimic the real world data. We investigate the use of different blending techniques. The outcome of our study demonstrates that synthetic images, when integrated with an existing real training set, substantially improve the object recognition accuracy of the model without compromising inference time. Our code can be found at https://github.com/EduardBartolovic/synthetic-data-generation