This project allows a user to search within the frames of a video for the nearest object. A user selects the item or region of interests, which corresponds to a bag-of-words corresponding to the item. Items are compared using SIFT (Scale Invariant Feature Transform) descriptors.
The code uses a visual vocabulary of 90,000 words using a random sample of 15 frames from each provided frame. Using the SIFT descriptor of a given frame, we create a histogram that model the distance of each descriptor to a visual word. These histograms can then be compared against one another to find similar distances. We use the Euclidean distance and k-Means to best associate features. This program was also optimized using tf-idf weighing (removing common words and using a matrix of weights to better associate frames).
This project is able to cut down the width and height of an image by removing low intensity pixels. This is achieved through a dynamic programming approach that analyzes similarity between nearby pixels, and cutting the optimal collection of pixels from an image.
This program is able to combine multiple images using a homography matrix computed between the two images.
