This thesis revisits a novel statistical approach to image feature, specifically line and edge detection, using the covariance model based image feature detector. This research explores the implementation of covariance model-based image feature detection and any of its intermediate results in combination with deep learning approaches. The result shows that incorporating such prior information is useful in making more robust, non-spurious detections.
Moreover, in the exploration of the CVM operator, a feature descriptor was designed that makes use of the CVM convolution kernels. This descriptor shows a rotation and limited scale invariant capability, which has shown to be useful for keypoint matching.
To join the presentation via Microsoft Teams click here