Dataset simulation method and system for multi-view three-dimensional reconstruction

Abstract

The application discloses a kind of data set simulation methods and systems for multi-view three-dimensional reconstruction, method includes the following steps: in simulation environment, configure multiple virtual sensors, and the data of multiple virtual sensors is synchronously collected;The global geometric information of virtual camera is used to calculate the view cone overlap rate of image, and the candidate image pair with effective view cone overlap rate is initially screened out;For any candidate image pair, based on geometric back projection and re-projection, the dense correspondence relationship at pixel level is calculated, and using depth consistency check, the matching true value pair without noise is generated by eliminating occlusion area;According to the geometric indexes of effective view cone overlap rate, observation distance ratio and view angle difference, the matching true value pair is quantitatively evaluated and difficulty classification is carried out according to the quantitative evaluation result, and the standardized data set containing dense corresponding true value, relative pose and difficulty label is packaged and output.The application realizes the low-cost large-scale generation of standardized data set without noise and full-pixel matching true value.

Images