Name: LEONARDO DE ASSIS SILVA
Publication date: 16/12/2016
Advisor:
Name |
Role |
|---|---|
| RAQUEL FRIZERA VASSALLO | Advisor * |
Examining board:
| Name |
Role |
|---|---|
| EVANDRO OTTONI TEATINI SALLES | Internal Examiner * |
| MARIANA RAMPINELLI FERNANDES | External Examiner * |
| RAQUEL FRIZERA VASSALLO | Advisor * |
Summary: Considering the increase in device processing capabilities, images can be used to analyze scenes and extract three-dimensional information from pixels. The process of retrieving the three-dimensional information from the environment is called 3D reconstruction. Estimating precisily the depth of homogeneous regions in an image is still a challenge in computer vision. In this work, two dense 3D reconstruction methods were proposed, aproximating homogeneous regions by planes and by a surface obtained from Delaunay triangulation. Our approach aims to achieve a good trade-off between precision and processing time. It is worth to mention that this work is focused only on the reconstruction stage. The intrinsic parameters are considered to be obtained by calibration, while the extrinsic parameters are estimated from a camera tracking process. The reconstruction estimation process needs a set of 10 images that presents some overlap with a reference image. The matching for every pixel in the reference image is searched in all the 10 images, resulting in a sparse estimation after a fusion and filtering stage. It is possible to decrease the algorithm processing time by using a gaussian multiresolution pyramid. The homogeneous regions are identified with SRM technique and their depth estimation is obtained from the sparse points cloud reconstructed a priori. This is done using one of the methods propoused: the planar aproximation or the surface obtained through Delaunay triangulation. The results were satisfactory according to the criteria of precision, recall and processing time. Comparing with DTAM, a real-time reconstruction technique, the results achieved better accuracy, especially in homogeneous regions. The longer processing time is due to the lower capacity hardware when compared to the hardware used for the DTAM experiments. Additionally, the implemented code is not optimized. In this way, it is possible to obtain better reconstruction results with the proposed method. Considering the evaluations carried out in this work, the obtained results may be considered promising.
