🧞 Requesting feedback and suggestions

Hi all,

We are constantly trying to make Food Recognition Benchmark better for everyone and would really appreciate your feedback. :raised_hands:

  • This iteration features a newly designed starter kit, so you no longer need to worry about random environment variables. :wink:
    But did it help or you are missing some resources from the older format?

  • The number of classes and images has grown over time as well, and it is just going to increase in future rounds! :chart_with_upwards_trend:

Feel free to reply to this thread with your suggestions and feedback on making the competition better for you!

  • What have been your major pain points so far?
  • What would you like to see improved before Round 2?
  • Do you need additional resources from us? if yes, on which topic/library?



Hey @shivam,

I have a few questions:

  1. Are we merging some of the classes in Round 2?
  2. Is there a way to skip the large images? I do that in augmentation but it affects the performance somehow.

Hi @gaurav_singhal,

  1. Yes, we are working on it and this should be the case with Round 2 mostly.

  2. You ideally shouldn’t need to worry about it, with the new timeout logic.
    BUT if you still want to handle it manually:

    • You can read image metadata using PIL/exif, etc in your submission and ignore those files.

      NOTE: This process will give you accurate file dimensions, etc but don’t try to load images again and again, because it might reduce your performance during prediction. In case you are loading an image, you can potentially send it across as an image (and not only filename) to inference. You might need to check MMdetection’s interfaces for it.

    • Another approach that might not be 100% reliable but way faster – check the file’s size using os. stat or so, and get a rough estimate about the image dimensions, and skip it accordingly.