Alignment of Images Captured Under Different Light Directions
Abstract:
Image alignment is one of the first steps for most computer vision and image processing algorithms. Image fusion,
image mosaicing, creation of panoramas, object recognition/detection, photometric stereo and enhanced rendering
are some of the examples in which image alignment is a crucial step. In this work, we focus on alignment
of high-resolution images taken with a fixed camera under different light directions. Although the camera position
is largely fixed, there might be some misalignment due to perturbations to the camera or to the object, or
the effect of optical image stabilization, especially in long photo shoots. Based on our experiments, we observe
that feature-based techniques outperform pixel-based ones for this application. We found that SIFT [Low04] and
SURF [BTVG06] provided very reliable features for most cases. For feature-based approaches, one of the main
problems is the elimination of outliers, and we solve this problem using the RANSAC framework. Furthermore, we
propose a method to automatically detect the transformation model between images. The datasets that we focus
on have around 10-100 images, of the same scene, and in order to take advantage of having many images, we
explore a graph-based approach to find the strongest connectivities between images. Finally, we demonstrate that
our alignment algorithm improves the results of photometric stereo by showing normal maps before and after
alignment.