Regions from the MSER and MSCR detectors come in a wide range of shapes, and are less affected by background changes than the texture patches normally used to describe them in recognition and matching applications. This means that descriptors computed directly from the regions could also have better invariance properties. We demonstrate that while texture descriptors still are superior on planar scenes, shape descriptors offer distinct advantages in full 3D scenes, and under large illumination changes. In general, a system which makes use of both shape and texture is to be preferred, since this will work under a wider range of scene categories.
|
|
| Input image | 64 randomly selected MSERs |
|
|
| Affine normalized intensity patches | Affine normalized shape patches (new) |
The shape descriptor algorithm was presented at the ICCV'07 conference.