Image Similarity Challenge - one month until Phase 2!

Greetings, Facebook Image Similarity Challenge participants!

To everyone participating in the Image Similarity Challenge, thanks for all your great work so far!

As a reminder, Phase 2 of the challenge starts in around a month and will run for 48 hours. In this phase, you will have the opportunity to make up to three submissions, and performance will be used to determine prizes.

In Phase 2, you will receive a new, unseen query set of 50,000 images and make a new submission using the model you have built in Phase 1. Manual annotation of the Phase 2 query set or re-training of your model is prohibited and will result in disqualification. Prize-eligible solutions must also treat each unseen query image as an independent observation (i.e., not use any information from other query images) when producing their submission. As with Phase 1, the reference images should also be treated independently.

Phase 2 of the competition will take place from October 26, 2021 00:00 UTC to October 27, 2021 23:59 UTC. If you would like to be eligible for final prizes, please ensure in advance that your team will be available during this period to download the new data and run your model.

As this final phase approaches, here are a few things to keep in mind:

  • Phase 2 dataset: The Phase 2 dataset will be a new set of 50,000 unlabeled query images, where some query images are derived from the reference set and the remainder are distractors. Keep in mind that, as discussed in the challenge paper, the Phase 2 query images will include a few transformations not seen in the Phase 1 dataset. You’ll want to make your solutions as robust as possible to new transformations.
  • Rules: Make sure that you are following the competition Official Rules and adhering to the Rules on Data Use. We will not hesitate to disqualify non-conforming submissions. If you have questions, you can also review the discussion forum and then ask for clarification if something remains unclear.
  • Freeze and submit your code: In generating your Phase 2 submission, there can be no changes in your code aside from reading data from a new Phase 2 query dataset. Before October 19, 2021 at 23:59:59 UTC, you must submit a copy of your code as described on the “End Phase 1 Submission” pages for the Matching and Descriptor tracks (you will need to be logged in to access these pages). Note that for Descriptor Track participants, you also need to submit a file with your final version of the reference descriptors.
  • Participate! Don’t be discouraged if you’re not near the top of the leaderboard. Remember that the Phase 1 leaderboard includes scores for provided labels, so it doesn’t necessarily give a complete picture of how well other participants’ models will generalize to new data. You may be doing better than you think.

Thank you once again to everyone who has participated in this competition. We look forward to seeing what you can do in this final stretch!

The DrivenData Team

Hi,@dd-mike,
We want to know whether we can download the data with a password before phase 2 starts, so that we can run our model directly at October 26, 00:00 UTC. Downloading costs us a lot of time.

< this is deleted message>

Hi qqerret. Unfortunately, that’s not possible. All participants will receive access to public S3 buckets at the same time at October 26, 00:00 UTC. Note that you will only need to download a new query set (~7GB for Phase 1), since the reference set will remain the same.