Search for a command to run...
To ensure good performance, modern machine learning models typically require large amounts of quality annotated data. Meanwhile, the data collection and annotation processes are usually performed manually, and consume a lot of time and resources. The quality and representativeness of curated data for a given task is usually dictated by the natural availability of clean data in the particular domain as well as the level of expertise of developers involved. In many real-world application settings it is often not feasible to obtain sufficient training data. Currently, data augmentation is the most effective way of alleviating this problem. The main goal of data augmentation is to increase the volume, quality and diversity of training data. This paper presents an extensive and thorough review of data augmentation methods applicable in computer vision domains. The focus is on more recent and advanced data augmentation techniques. The surveyed methods include deeply learned augmentation strategies as well as feature-level and meta-learning-based data augmentation techniques. Data synthesis approaches based on realistic 3D graphics modeling, neural rendering, and generative adversarial networks are also covered. Different from previous surveys, we cover a more extensive array of modern techniques and applications. We also compare the performance of several state-of-the-art augmentation methods and present a rigorous discussion of the effectiveness of various techniques in different scenarios of use based on performance results on different datasets and tasks.