View Invariant Point Matching with Applications to Object Recognition
Saved in:
| Title: | View Invariant Point Matching with Applications to Object Recognition |
|---|---|
| Authors: | Lotfian, Sina |
| Committee Members: | Foroosh, Hassan |
| Summary: | Estimating the geometry relation between two images is one of the key steps for various computer vision tasks including visual tracking, 3D reconstruction from multiple images, optical flow estimation, and camera calibration. The error in the estimation phase can propagate into final accuracy of the aforementioned task, therefore reducing the estimation error has important practical implications. The first step in geometry transformation is establishing point correspondence between two images. Even though, the robust estimators can prune out some outliers later in the calculation pipeline, higher matching accuracy leads to better estimate. In this thesis, we demonstrate how use of multiple level of refinement in multiple scales can increase the accuracy of keypoint matching. Next, we show that the invariant geometry of matches can help us match the images between two images of the same object under different view points. The view invariant parameter that we examine in this thesis is the difference between eigenvalues of homography matrix induces by triplets of points between images. Every triplet of points in the image induce a homography matrix. In the special case of two cameras viewing the same scene under different view points, the homography is reduced to the special case of homology with two equal eigenvalues. This enables us to define a error measurement for the task of template matching with application in view-invariant object retrieval and face recognition. Finally, we discuss the possible attacks on the pipeline of estimating transformation between two images. We propose a method that targets the extraction and matching phases of the pipeline in order to derail the calculation process. Small perturbation is applied on the image patches so that the output geometry matrix is wrong. The magnitude of distortion in the images are kept limited so that human eyes are unable to distinguish the perturbed image. The attack can be conducted on either one (single image attack) or both images (dual image attack). |
| URL: | https://stars.library.ucf.edu/etd2020/1919 |
| Database: | OpenDissertations |
| Abstract: | Estimating the geometry relation between two images is one of the key steps for various computer vision tasks including visual tracking, 3D reconstruction from multiple images, optical flow estimation, and camera calibration. The error in the estimation phase can propagate into final accuracy of the aforementioned task, therefore reducing the estimation error has important practical implications. The first step in geometry transformation is establishing point correspondence between two images. Even though, the robust estimators can prune out some outliers later in the calculation pipeline, higher matching accuracy leads to better estimate. In this thesis, we demonstrate how use of multiple level of refinement in multiple scales can increase the accuracy of keypoint matching. Next, we show that the invariant geometry of matches can help us match the images between two images of the same object under different view points. The view invariant parameter that we examine in this thesis is the difference between eigenvalues of homography matrix induces by triplets of points between images. Every triplet of points in the image induce a homography matrix. In the special case of two cameras viewing the same scene under different view points, the homography is reduced to the special case of homology with two equal eigenvalues. This enables us to define a error measurement for the task of template matching with application in view-invariant object retrieval and face recognition. Finally, we discuss the possible attacks on the pipeline of estimating transformation between two images. We propose a method that targets the extraction and matching phases of the pipeline in order to derail the calculation process. Small perturbation is applied on the image patches so that the output geometry matrix is wrong. The magnitude of distortion in the images are kept limited so that human eyes are unable to distinguish the perturbed image. The attack can be conducted on either one (single image attack) or both images (dual image attack). |
|---|
