Crowdsourcing at ML Boot Camp. We consider mIOU without pictures for a new task from Odnoklassniki

    Hello! Hot summer. The organizers of the "IT" championships sat a lot in the sun, burned and grabbed a blow, but most importantly, they gathered a new task for the next (already ninth) contest on the ML Boot Camp platform. The championship will be held online within a month.

    Now more than 10 thousand specialists are registered on the platform. It often happens that the tasks that appear on it are not for everyone's teeth (including the Kaggle gods). Why are we doing this? It is necessary to develop and try, moreover, on real data, and not a synthetic bullet. Victories will come with time.

    The task that we want to offer you to solve in this competition is different from all the previous ones. The theme of the task is the detection of objects in images. The statement of the problem implies that there will be pictures in the data set, but, funny enough, they are not. And these are not petabytes of data. And not even gigabytes.

    The topic is popular and, in fact, very important. In Odnoklassniki social network, as in other products, there is a task to filter content. Imagine that your child will see the wrong picture - his world will forever change.

    Two years ago, Alexey Sennikov wrote an article about how Odnoklassniki attract their users to solve this problem. Briefly, in an article, Alexey talked about the Odnoklassniki Moderator game application , in which users of a social network classify images as good and bad, receiving various buns for it, and also shared a way to solve the problem of increasing DAU (Daily Active Users) applications using algorithms machine learning.

    By the way, DAU now has 40 thousand users, and more than 1'000'000 decisions per day. Not bad, huh?

    The challenge of the new contest

    Odnoklassniki social network has a platform for marking up data. On it, users were given the task of highlighting a given object in a photograph with a rectangle. According to the answers of people, it is necessary to restore the true position of the object.

    At the entrance you will receive crowdsourcing markup and data in the form of all your favorite labels. In response, you will just have to send a file with predictions.

    The dataset and base line will be published on the day the competition starts at the ML Boot Camp.

    To evaluate the solution, the mIOU (mean intersection over union) metric will be used. If you have not encountered this metric, we recommend that you read an article about it .


    The championship will go online. We start on June 27 at 19:00, close the submissions on July 29 at 12:00. We are also summing up the results on July 29 at 13:00.

    All registered users on the platform will receive a reminder in the mail. Sign up or subscribe to the @mrgchamps channel . Along the way, enter the community of participants ( chat @mlbootcamp for 1400 people) in Telegram to be exactly in the subject of everything that happens.


    Where without them. Firstly, the best participants will receive:

    1st place: MacBook Pro 13 ', 2-core processor, 256 GB SSD, 16 RAM.
    2nd place: iPhone XS Max, 256 GB.
    3rd - 6th place: Apple Watch Series 3 42mm or Samsung Gear S3 Frontier to choose from.
    7th - 10th place: Western Digital My Passport 4 TB.

    Secondly, cool t-shirts will get the top 30% of the total number of participants.


    For a good start in the championship, we recommend reading the articles:

    1. Grouper: Optimizing Crowdsourced Face Annotations
    2. Maximum Likelihood Estimation of Observer Error-Rates Using the EM Algorithm

    Come, participate, learn and win!

    Also popular now: