1. Become a Developer for CrazyEngineers - Apply NOW!

UC Berkeley Researchers Create Tool To Average Images

Discussion in 'Computer Science | IT | Networking' started by Chaitanya Kukde, Aug 26, 2014.

  • by Chaitanya Kukde, Aug 26, 2014 at 9:19 PM
  • Chaitanya Kukde

    Chaitanya Kukde Maniac

    Engineering Discipline:
    Electronics & Communications
    The amount of visual data is growing and it is growing fast. How fast? See for yourself:
    • Facebook users upload 350 million new photos each day (Business Insider)
    • Snapchat sees even more, about 400 million photos each day (Techcrunch)
    • 27,800 photos are uploaded to Instagram every minute (Popphoto)
    • Google Maps alone has over 20 petabytes (about 20.5k TB) data (Mashable)
    • Jan Koum of WhatsApp claimed over 400 million photos per day (Twitter)
    Now all that data needs sorting, storage, metadata and other things the Big Data guys know better about. The scientists at UC Berkeley may have found a solution to tame this monster. Their solution is simple: Average the data. Scientists Alexei Effros, Jun-Yan Zhu and Yong Jae Lee developed a software tool to develop an image representing an average of a cluster of images. This tool, named AverageExplorer was presented as a part of paper at SIGGRAPH 2014 conference in Vancouver, Canada.

    AverageExplorer makes it easier to analyse 'chains of related images' such as that of brides and grooms at weddings or that of a graduation ceremony. Specific features in images could be given more 'weight' by the user to create subcategories and make sorting easier. The researchers said that potential applications include online shopping, media analysis, sports analysis, etc.

    image_average_1.
    The tool lets users focus on images in which President Obama appears over Stephen Colbert’s shoulder. Users can also observe Colbert’s typical body posture among those results.
    Yong Jae Lee further stated that the efficiency of Computer Vision systems in deducing key features from images could be improved using this technique. For example, a face recognition system needs to be trained using lot of data which requires a considerable manual effort. Instead, if the average image is annotated for features (such as eyes, nose and mouth), the effect could propagate down the collection to each individual image.

    It is worth to note here that an artist, Jason Salavon has created average images from hundreds of photos of kids with Santa, newlyweds or baseball players to illustrate this concept of 'image averaging'. Check out his amalgamation work on Chicago city's photos taken from the city's typical tourist points here. But unlike Salavon, the scientists automated the process of aligning and sorting the images.

    On a parting note, if we could create an average of a number of averages, we could have an 'image within an image' or an "Inception"-styled image. We could call it 'Im'ception.

    Nevertheless, share your views in the comments. Also feel free to bash the author if you thought the 'Im'ception joke was a bad one.

    Source: UC Berkeley News
     
    • Like Like x 3
    Last edited by a moderator: Aug 26, 2014
  • Categories:

Comments

Discussion in 'Computer Science | IT | Networking' started by Chaitanya Kukde, Aug 26, 2014.

Share This Page