Seminar

Ing. Michal Kottman

Improving binary feature descriptors using spatial structure

Feature descriptors are used to describe salient image keypoints in a way that allows easy matching between different features. Modern binary descriptors use bit vectors to store this information. These descriptors use simple comparisons between different keypoint parts to construct the bit vector. What they differ in is the arrangement of keypoint parts, ranging from random selection in BRIEF descriptor to human vision-like pattern in the FREAK descriptor. A recent descriptor D-Nets shows that line-based arrangement improves recognition rate for feature matching. We show that by extending the comparisons to line arrangement better describes the spatial structure surrounding the keypoint and performs better in standard feature description benchmarks.

Ing. Andrej Fogelton

Superpixel Image Clustering

Superpixel image preprocessing brings considerable speedup to object recognition algorithms. The purpose of these algorithms is to cluster pixels based on color and location information. These homogenious clusters are in similar predefined size. Color and location coherence in most cases indicates the association to the particular object. Such an information is very usefull in object recognition algorithms, whos recognition process run per pixel. Whith these kind of information significant speed up can be achieved to adjust the algorithm to run on per superpixel bases. We present and evaluate several experiments on Simple Linear iterative Clustering (SLIC) algorithm, which is the fastest and best quality (boundary adherence) superpixel algorithm. Depending on the speed vs. quality requirements we provide such modifications.

kde: zasadačka UAPI – 4.posch. 4.08