Learning transformation invariance for pairwise image matching
Image matching is a fundamental problem in computer vision. In this thesis, we address the image matching problem as learning and classifying correspondences. More precisely, we formulate the image matching problem as: given a set of training pairs of images that implicitly captures the tra...
Saved in:
| 主要作者: | |
|---|---|
| 其他作者: | |
| 格式: | Theses and Dissertations |
| 語言: | English |
| 出版: |
2010
|
| 主題: | |
| 在線閱讀: | https://hdl.handle.net/10356/41433 |
| 標簽: |
添加標簽
沒有標簽, 成為第一個標記此記錄!
|
| 總結: | Image matching is a fundamental problem in computer vision. In this thesis, we
address the image matching problem as learning and classifying correspondences.
More precisely, we formulate the image matching problem as: given a set of training
pairs of images that implicitly captures the transformation(with both positive
and negative classes), identify if a new pair of test images is matched via the transformation
class. In this formulation, all the training data, as well as test data, are
image pairs. The approach taken is to consider only relative visual content, rather
than absolute visual content, so the learned image matching classifier could be applied
to images of totally different visual content as compared to the training data.
This is in contrast to appearance-based object detection methods, for which once
the training process is completed, the classifiers may only be used to recognize
objects of the same categories with the training images. |
|---|