Reconstruction of a 3D model is typically obtained using depth sensors or a stereo camera system. Such solutions are less accessible, therefore it is an interesting challenge in the field of computer vision to build a 3D model without additional equipment besides a camera. Our work discusses reconstruction based on images. First we examine the theoretical background and algorithms, then we proceed with description of our reconstruction system. We use a method called structure from motion to obtain object's sparse reconstruction and camera's extrinsic parameters. The latter are used for building a dense 3D model. Evaluation of extrinsic parameters shows that adding cameras to the reconstruction leads to accumulation of the error. Despite this occurrence the average error of the calculated poses remains within a few millimetres. Finally we evaluate the accuracy of the reconstructed models by comparing them with their ground truth. Results indicate small average errors, therefore we can conclude that the implemented system works sufficiently well.
|
