Method for Automated Data Collection for 3D Reconstruction

Abstract

The paper presents an automated method to guide a human operator during RGB image shot for improving the quality of further 3D reconstruction for low-rise outdoor objects and a use case for method application. The method provides automation as a three-step process: local analysis of images performed during the shooting, global analysis, and recommendations performed after the shooting. Method steps filter out defective images, approximate future 3D-model using Structure-from-Motion (SfM), and map it to human operator trajectory estimation to identify object areas that will have low resolution on the final 3D model, requiring additional shooting. Method structure does not require any sensors except RGB camera and inertial sensors and does not rely on any external backend, which lowers the hardware requirement. The authors implemented the method and use-case as an Android application. The method was evaluated by experiments on outdoor datasets created for this study. The evaluation shows that the local analysis stage is fast enough to perform during the shooting process and that the local analysis stage improves the quality of the final 3D model.