Computer Vision Visual Search and Report Coursework 2019
This report reviews the main techniques used in a basic visual search system, including the various descriptors that can be obtained from images and several methods to compute the similarity of images to produce a sorted image collection. This is then followed by an in-depth discussion of experimental results of a developed program (written in MATLAB) to visual search a dataset of 591 images in 21 categories developed by Microsoft MSVC-v2 available at http://download.microsoft.com/download/3/3/9/339D8A24-47D7-412F-A1E1A415BC48A15/msrc_objcategimagedatabase_v2.zip.
A major challenge in this project was from the supplied MSVC dataset. This is due to the nature of the categorisation of the images and that some images were repeated, and some objects/features inside the images were included in more than one category. This meant that some of the results were not a true representation of a visual search.
In conclusion, a visual search system has a variety of contributions determining its accuracy and performance. These contributors can involve different descriptor techniques using quantisation and gridding, the dimensionality of the feature space and types of distance measures to decide on similarity.