Voogle: A Visual Search Engine

We detail the design and implementation of Voogle, a content-based video search engine that takes as input a short video clip and outputs other videos of similar visual content. We leveraged Dense Trajectories for feature extraction, quantize the features using a bag of visual words model, and applied the k-nearest neighbor algorithm using Euclidean distance as a metric for similarity. We scraped approximately 20,000 videos from Vine, a mobile application that allows users to upload short 6-second videos. Results indicate that feature extraction dominates execution time comprising 94.1% of overall execution time for the full end-to-end pipeline (when including offline and online portions). Furthermore, storage of features is identified as a core problem as storing the features consumes three orders of magnitude more storage space than the actual video (.mp4) itself. Sample query results of the end-to-end system can be viewed at http://homes.cs. washington.edu/ ̃cdel/voogle/landing.php (use Safari v8.0.6+).