| 研究実績の概要 |
a)For removing photometric outliers in the texturing of multi-view stereo reconstructions, we looked into mathematical techniques that perform rank reduction on matrices of observations with high data redundancy while also detecting outliers. We have generated some initial datasets, both synthetically rendered and real-world multi-view stereo datasets that suffer from the outlier problems we try to tackle. Currently we are looking into methods that jointly perform rank reduction and image alignment since misalignment influences outlier removal and vice versa.
b)We developed a method for calibrating point light source positions, which is required for image-based modeling techniques such as photometric stereo. Previous works on this are expensive to set up and tedious to use. Our method works with needle pins that cast shadows on a flat board. Given shadow observations under enough board poses, we can jointly estimate light and needle pin positions. The work is scientifically interesting because it shows a mathematical connection between point lights, pinhole cameras and structure from motion.
c)WeI started collaborating in a project where we analyze the auto-recalibration of stereo fisheye cameras. This is highly relevant, e.g., in autonomous driving or consumer products, where fisheye lenses provide better overview over the scene. It seems that, in contrast to the well-studied pinhole and radial distortion camera models, fisheye models do not exhibit a mathematical ambiguity between camera intrinsics and extrinsics during the calibration.
|
| 今後の研究の推進方策 |
We continue preparing datasets for the 3D reconstruction texturing outlier removal, in particular real-world datasets of various sizes where our investigated effects (outliers + misalignment) occur to a strong degree.
Then I want to implement two solution methods, RASL (a method based on matrix rank reduction and simultaneous image alignment) and an own solution based on optical flow and robust homography fitting, to compare them in terms of 1) their solution quality and 2) their sensitivity to parameter choice. The problem I experienced so far with low-rank approximation methods is that parameter configurations working in one scene part often do perform poorly in other scene parts and it is not obvious how to choose the parameters appropriately.
The goal is to submit this work to the Conference on Computer Vision and Pattern Recognition in November.
|
