: Recent years have witnessed a rapid development of machine learning theories and deep learning algorithms in the field of computer vision, which have been widely applied in task scenarios such as object detection, semantic segmentation and action recognition. However, model perform ances in practical deployment always rely on the assumption that the training and testing data are iden tically and independently distributed, and prone to be affected by domain shift. Domain shift, namely distribution shift between the training domain and the test domain, poses a great challenge to the mod el generalization, thereby making domain generalization an important research topic in computer vi sion. Domain generalization aims to train a model on a single or multiple domains and maintain good generalization on unknown target domains with different distributions, which provides important guar antees for model deployments. This paper clearly elaborates the domain generalization technique in the computer vision field and its research progress through combing and summarizing representative resear ches over the past decade. Firstly, the paper makes a comprehensive introduction of domain generali zation from three aspects, i. e. , task formulation, task characteristics and research ideas. Then, the paper catergorizes existing approaches into three groups according to the main research ideas to deal with domain shift, namely, data augmentation, model optimization, and domain gap mitigation. Sub sequently, the applications of domain generalization techniques in computer vision and related large scale public datasets are introduced. Finally, the paper discusses problems needed to be solved in the current domain generalization field and possible future research direction.