Visual Wikipedia: Exploring Paris & ! Aayush Bansal, Akanksha Saran, Krishna Kumar Singh, Pranay Agarwal, Priyam Parashar {aayushb,asaran,ksingh1,pranaya,pparasha}@andrew.cmu.edu

Why is this city so awesome? Oooh! Where do I go? Where are the cool places? LUCKY YOU! YOU ARE ASKING THE RIGHT ROBOT. LET ME JUST RUN MY ROUTINE ON Images were divided as per the city blocks shown and were Discriminative clustering of patches High Score Low Score THIS CITY'S extracted using Google Street View API IMAGE DATASET. Exploring Paris A Tour of Pittsburgh

David McCullough Cathedral Of Learning

Schenley Drive Heinz Field

Approach: Discriminative Clustering (a) Rue de Rivoli is one of the most famous streets of Paris, a commercial (a) Heinz Field is a stadium located in the neighborhood of Pittsburgh, street whose shops include the most fashionable names in the world. Pennsylvania. It primarily serves as the home to the Pittsburgh Steelers and University of Pittsburgh Panthers American football teams. Standard Clustering (K-means) techniques do not reject outliers. (b) Just outside Jardin du Luxembourg is the Rue Auguste Comte and Avenue de l’Observatoire. A great example of the architecture at the end (b) The David McCullough Bridge is commonly and historically known as the 16th of the 19th century. Street Bridge. It is a through arch bridge that spans the in Pittsburgh, Pennsylvania. Discriminative Clustering: rejects outliers and learns the clusters which have maximum (c) Pont de la Tournelle (Tournelle Bridge in English), is an arch bridge inter-cluster distance and minimum intra-cluster distance. spanning the river Seine in Paris. It is classified as a historical monument. (c) The Cathedral of Learning, a Pittsburgh landmark listed in the National Register of Historic Places, is the centerpiece of the University of Pittsburgh's main campus in the (d) The Palais-Royal (French pronunciation: [pa.l wa.jal]), originally called Oakland neighborhood of Pittsburgh, Pennsylvania, United States. the Palais-Cardinal, is a palace located in the 1st arrondissement of Paris. Garden-side view with the columns of the former Galerie d’Orlans. (d) Schenley Drive is a green stretch which runs along-side attractions like Schenley Discriminative K-means Approach Used Plaza, , etc.

Algorithm Algorithm Comparison Nearest-Neighbour

Partition data into D1 and D2. Initialize That's cool right! Initialize using K-means. I just came from I wonder if there are any Represented here are top using K-means on D1. Pittsburgh and images of Paris using the you just visited Paris similarities... For each cluster, train a discriminative Pittsburgh SIFT based BoW model. Prune the clusters having less than 5 The nearest neighbor in classifier treating other clusters as Paris images was found samples. Paris using histogram negative intersection. If nearest data. neighbors are also one of For each cluster, train a discriminative the high scoring images Similar Patches classifier treating other clusters as negative between Pittsburgh and Paris of Pittsburgh, then they Reassign patches to cluster with highest set and correctly-classified data as positive are considered as a score. positive matches set. Approach otherwise classified as negative. Repeat Cluster D2 using above discriminative Positive Matches Negative Matches classifiers.

Ignore the ones having confidence score I know!!! Iconic Images less than 0.25. Prune the clusters having Discriminative clustering of patches less than 5 samples Downloaded top 300 images for the resultant high scoring places from Swap D1 and D2. Flickr. Represented Future Work images using GIST feature References and implemented K-means Repeat 1) Improving nearest neighbor algorithm. clustering on them. [1] S. Singh, A. Gupta and A. A. Efros, Mid Level 2) Trying different clustering techniques, and for different image Discriminative Patches, In European Conference

features. on Computer Vision (2012). Top Picture: Shows clustering done for images 3) Combining user review (text based) for different locations with extracted for Heinz Field [2] C. Doersch, S. Singh, A. Gupta, J. Sivic, and Bottom Picture: Shows clustering done for our current score of visual uniqueness. A. A. Efros, “What Makes Paris Look like Paris?” images of Cathedral of learning As can be seen, one of the clusters shows ACM Transactions on Graphics (SIGGRAPH some meaningful relation like night-view of 4) Extending our approach to other popular cities and finding 2012), August 2012, vol. 31, No. 3. Heinz Field or portrait shot of Cathedral of similarity pattern. Learning, while the other cluster is a collection of random images or interiors specifically