Or use documents with high/low ratios of pos/neg words to automatically assign labels

Placing task Organized by Pavel Serdyukov Vanessa Murdock

Location-aware Web 2.0 apps are on the rise GPS became really pervasive technology Location-aware web services

Photos gets geo-tagged Last available stats on Feb, 2009: 100 million out of 3 billion Flickr photos are geotagged But that’s still around 3-4 % of all Flickr photos…

Photo sharing sites become “Geo” Yahoo! (Flickr) and Google (Panoramio) are main players

What about videos? Youtube example

Flickr Videos: “It’s like a photo, but it moves!” … and many are geotagged!

Data requirements Requirements for the data: Should be geotagged Geotags should be true and accurate Videos should be Creative Commons licensed Sample should dispersed enough But to download, we need to know the links Flickr API is more than enough: Search API provides a lot Possible to search using bounding boxes Up to 4000 links to photos/videos per search query So, technically, you can get them all

Crawling strategy Start from large bounding boxes covering the world For minLongitude, minLatitude, maxLongitude, maxLatitude: 160,-90,170,-80 0 photos 160,-80,170,-70 1208 photos 160,-70,170,-60 0 photos 160,-60,170,-50 19 photos 160,-50,170,-40 16714 photos 160,-40,170,-30 6 photos 160,-30,170,-20 6175 photos 160,-20,170,-10 722 photos Divide large cells into bounding boxes: 1) Check inner cells with side 2.0, if photos > 4000 2) Check inner cells with side 0.5, if photos > 4000 3) Check inner cells with side 0.1, if photos >4000 Retrieve 4,000 from each of 4 last years So, very popular 10km X 10km cells – up to 16,000

Datasets: Video Found ALL CC-licensed geotagged videos All geotags are very accurate 16, maximum zoom “street” level ~17,000 in total ~60% from North America Downloaded using 1 PC / 3 months Not all videos are downloadable Only mp4 are directly linked, flv are embedded! Downloaded: 10,216 videos plus metadata

Video metadata: Example

Video metadata: Example

Datasets: Images Crawl plan: CC-licensed images with geotags Accuracy: 6 (region level) - 16 (street level) What we crawled in 4 days: Out of total ~23 million Geo-CC-photos (API counts) Located ~15 million: Including links to images, userids, timestamps, tags, geotags Sampled from them to process and share Sampling was really ad-hoc: All 1km grid cells with < 5 photos are removed Cells with 5 – 15 photos are left untouched Cells with > 15 photos: sampled 16% of photos Result: ~3,185,258 URLs/metadata to share

Videos vs. Photos Videos and Photos: what’s the difference? Videos are sets of “photos” with sound Advantages / disadvantages: More photos means more evidence Lots of these “photos” are noise Extracted frames and features: Saved every 4th second of a video as JPG Extracted 9 visual features: For ~100,000 frames For ~3,100,000 photos Sound would be cool, but hard to train models

Task description Shared as training data: 5125 videos/metadata/keyframes/visual features ~3,100,000 photos/metadata/visual features Shared as test data: 5091 videos/metadata/keyframes/visual features Only geotags are missing (removed by organizers) Test/training split: by user ids Task: predict geotags (so, two decimals) Ground truth: Ready-made by owners Evaluated: by number of videos placed within certain distance from the actual location 1, 5, 10, 50 and 100 km

Examples: Easy http://www.flickr.com/photos/63666148@N00/3615989115/

Examples: Hard

Results 5 groups submitted 21 runs: ~1600 more than next best ~300 more than next best