AbstractMatching points between images is difficult because of the many possible variations between images of the same scene or 3D feature.
This includes different viewpoints, changing lighting conditions, occlusion and noise.
In this work, I demonstrate that the use of a local descriptor which is individual to each interest-point improves matching performance over using one globally.
I propose two different approaches.
The first is to use a pooling operation, based on geometric blur, which is individual to each interest-point.
This is achieved by estimating how each interest-point will appear in other images through generating synthetic warps.
The second is to learn an optimal combination of base-descriptors for each interest-point so as to obtain an optimal descriptor for each interest-point.
This is achieved through supervised learning based on having multiple instances of the same interest-point.
Another difficult problem is detecting people in images of artwork, because of the huge variation in the ways people are depicted.
This included the media used (e.g. pencil, paint and sculpture) and the range of poses and projections, including Cubism at the extreme.
I demonstrate that state-of-the-art CNN based methods yield improved performance if the models are fine tuned on artwork, however their performance on photos is subsequently reduced.
This shows that these approaches cannot simultaneously generalise over and perform well on both photos and artwork.
The underlying theme of this dissertation is the proposition that these and other algorithms lack generalisation because their invariance is at too low a level, lacking high level semantic information.
For further work, I suggest the incorporation of high level semantic information into interest-point matching and better modelling of the structure of people which is shown to be essential for detecting people across different depictive styles.
|Date of Award||19 Jun 2019|
|Supervisor||James Laird (Supervisor), Neill Campbell (Supervisor), Peter Hall (Supervisor) & Matthew Brown (Supervisor)|