Generating dataset for object recognition from virtual world

Recent success in object recognition has been driven by deep convolution neural networks trained on large datasets. However, the construction of large datasets was extremely costly due to large amount of manpower required, and it is impossible to obtain real images to construct large datasets in some scenes. In this paper we present two approaches to generating synthetic images instead of real images to build datasets for object recognition, which requires little manpower in labeling and generating images. What is more, we studied the impact of factors in the dataset construction process on the quality of the dataset and compared two ways of generating datasets with testing in the real world. Our experiments evidence that the methods of constructing synthetic datasets can solve the object recognition problem to a certain extent, and we give the appropriate factors settings in the process of building synthetic datasets.