MIR2006

Uploaded from authorPOINTLite
Views:
 
     
 

Presentation Description

No description available.

Comments

Presentation Transcript

Generating Summaries and Visualization for Large Collections of Geo-referenced Photographs: 

Generating Summaries and Visualization for Large Collections of Geo-referenced Photographs Alexander Jaffe*, Mor Naaman*, Tamir Tassa†, Marc Davis$ *Yahoo! Research Berkeley †Open University of Israel $Yahoo! Research

Attraction Map of Paris: 

Attraction Map of Paris Stanley Milgram, 1976. Psychological Maps of Paris

Attraction Map of London: 

Attraction Map of London Jaffe et al, 2006.

Information Overload?: 

Information Overload? Flickr “geotagged”

Overview: 

Overview Problem definition Intuition for solution Algorithm for summarization Visualizing the dataset Evaluation Demo?

Problem Definition: 

Problem Definition Dataset: (photo_id, user_id, latitude, longitude) (photo_id, tag) Result: (photo_id, rank) Given all photos from a geographic region, find a “representative” summary set

Issues to Tackle: 

Issues to Tackle Noisy data Whatever, color, city, spectrum, santa barbara, california, usa, Lookatme, Herbert Bayer Chromatic Gate Photographer biases In locations In Tags Wrong data

Intuition: 

Intuition More “activity” in a certain location indicates importance of that location Tag that are unique to a certain location can suggest importance of that location

(Very) Simple Example: 

(Very) Simple Example

Algorithm Overview: 

Algorithm Overview Hierarchical Clustering of the location data For each cluster, generate cluster score Recursively generate ordering of all photos in each cluster, based on subcluster score and ordering

The Clustered Return of the (Very) Simple Example!: 

The Clustered Return of the (Very) Simple Example! 4, 6, 5 8,7 4,8,6,5,7

Generating a Summary: 

Generating a Summary A complete ranking is produced for all photos in the dataset An n-photo summary is simply the first n photos in this ranking.

Generating Cluster Scores: 

Generating Cluster Scores Main Factors: Number of photos Relevance (bias) factors “Tag Distinguishability” “Photographer Distinguishability”

Tag Distinguishability: 

Tag Distinguishability A measure of uniqueness of concepts represented in the cluster (“document”) TF/IDF based Compute frequency of each tag (TF) Compute (inverse) frequency of tag in the rest of the dataset (IDF) Aggregate TF/IDF over all tags in cluster using L2 norm Or, if you like formulas: Read the damn paper!

Summary of San Francisco: 

Summary of San Francisco

Progress Bar (almost done): 

Progress Bar (almost done) Problem definition Intuition for solution Algorithm for summarization Visualizing the dataset Evaluation Demo?

Tag Maps: 

Tag Maps Observation: The algorithm identifies “representative” locations The algorithm identifies unique, important tags

Tag Maps: 

Tag Maps

Tag Maps: 

Tag Maps

Ok, how do we evaluate this?: 

Ok, how do we evaluate this? Direct human-evaluation of algorithmic results Evaluated Tag Maps with various weighting options Compared summaries to 3 base conditions Compared chosen locations to top 15 locations selected by humans (Milgram-style)

Maybe we have time for a demo: 

Maybe we have time for a demo

Maybe we have time for Q’s: 

Maybe we have time for Q’s http://zonetag.research.yahoo.com (applied in prototype cameraphone app) http://blog.yahooresearchberkeley.com (more on this and other topics)