The challenge is based on images acquired by a KFM from samples collected by using several. Each FLOTAC/Mini-FLOTAC sample generates very high-resolution images, split in hundreds of overlapping patches, saved in jpeg or png format, resulting in thousands of images made available for the competition, unpublished and not provided to other people before. Each patch may or may not contain parasites’ eggs. For training data, participants are also provided with ROI (region of interest) files containing, for each image, the coordinates of the bounding box of parasites within it, as identified and segmented by an expert operator. Test data is provided without the ROI files. The use of third-parties publicly available datasets, from the same domain or from different ones, is allowed, but their use must be declared to the challenge organisers.
Training and test datasets belong to different acquisition samples (i.e. different FLOTAC and/or Mini-FLOTAC devices). In particular, the test will happen in two different moments:
during the first period of the competition (day 1 up to 1 day to the deadline), a test sample (images coming from a single FLOTAC and/or Mini-FLOTAC device) will be made available for participants to test the performance of their models;
on the last day, a new test dataset (with images coming from more than one FLOTAC and/or Mini-FLOTAC device) will be released. THIS WILL BE THE DATA USED TO DETERMINE THE WINNERS.
As the competition will be hosted on Kaggle, a subset of the test data will be used for the public leaderboard, while the remaining part will be only used for the private leaderboard. During the first period, 100% of the data will be used for the public leaderboard, while during the second period (the last day), only 5% of the data will be used for the public leaderboard. Teams will be allowed a maximum of 1 submission per day. This means that they will have a single submission on the last day to submit a prediction for the real test set.
The submission will consist of a .csv file where the participants have to report, on different lines and for each image, a pixel-encoded binary mask (please refer to the official Kaggle competition page for details) identifying all the detected eggs (if any).
Two are the available tracks (the first mandatory, the second optional):
Detection of parasites’ eggs. The F1 Score at different intersection over union (IoU) thresholds will be used as the performance metric (refer to the official competition Kaggle page). In case of a draw, the team with the lowest number of submissions made will be declared the winner
Also the inference time (on the test set, run on a HPC unix machine equipped with GPU acceleration) will be considered to declare the winner, according to the following formula:
score =2 * (F1IoU* normT) / (F2IoU + normT)
where normT is the min-max normalized execution time (median value over 10 repetitions) between all the participants
For both tracks the participants are asked to submit a Unix executable (or python code, as preferred). This will be used to reproduce the obtained results (both tracks) and to measure inference speed under the same workload conditions (track 2). Teams interested in taking part in Track 2 have to write their will on a topic on the discussion board.
We officially announce the "PARASITE-PAttern Recognition Applied SardInia TEam" to be the winner of the first AI-KFM 2022 challenge. The presentation of the challenge and the winning approach will be done on the 25th of May, during the ICIAP 2021 conference. The link to attend the event is the following
We thank again all the participants for being part of the AI-KFM 2022. Looking forward to seeing you in the next edition!
The AI-KFM Team