Local features are important building blocks for many computer vision algorithms such as visual object alignment, object recognition, and content-based image retrieval. Local features are extracted from an image by a local feature detector and then the detected features are encoded using a local feature descriptor. The resulting features based on the descriptors, such as histograms or binary strings, are used in matching to find similar features between objects in images.In this thesis, we deal with two research problem in the context of local features for object detection: we extend the original local feature detector and descriptor performance benchmarks from the wide baseline setting to the intra-class matching; and propose local features for consumer video scene boundary detection.In the intra-class matching, the visual appearance of objects semantic class can be very different (e.g., Harley Davidson and Scooter in the same motorbike class) and making the task more difficult than wide baseline matching. The performance of different local feature detectors and descriptors are evaluated over three different image databases and results for more advance analysis are reported.In the second part of the thesis, we study the use of Bag-of-Words (BoW) in the video scene boundary detection. In literature there have been several approaches to the task exploiting the local features, but based on the author’s knowledge, none of them are practical in an online processing of user videos. We introduce an online BoW based scene boundary detector using a dynamic codebook, study the optimal parameters for the detector and compare our method to the existing methods. Precision and recall curves are used as a performance metric.The goal of this thesis is to find the best local feature detector and descriptor for intra-class matching and develop a novel scene boundary detection method for online applications.
展开▼
