The bounding boxes in the submitted results overlapping the “others” objects with the threshold larger than the threshold or locating in the ignored regions will be filtered out in the evaluation.
NO. If you would like to report results of your algorithm with different settings (e.g., different parameters and training conditions), please use the training and validation sets for this purpose, and only submit one result to the evaluation server.
Similar to the PASCAL VOC dataset, we labeled some “ignore” regions, which indicates the regions containing the objects that are difficult to be annotated due to low resolution or crowd. Meanwhile, some rarely occurring objects (e.g., machineshop truck, forklift truck, and tanker) are labeled as “others”, which are ignored in evaluation.
Thank you! We are aware of some deficiencies in the provided annotations and are still trying to improve these. We do appreciate all kinds of feedback, so please don’t hesitate to contact us to report any findings.
Due to the academic purpose of the released dataset, we require the participators to use the academic email or the company email to register the account. We do NOT allow the participators to use the commercial email address (e.g., gmail, hotmail, yahoo, outlook, live, qq, 163, and 126) to register the account. If your do NOT have an academic or company email, please contact us. Meanwhile, if you face any other registration failure, please also send us an email. Notably, registering multiple account with different email addresses of the same team violates the challenge policy, which may lead to a permanent ban.
Based on the policy of the challenge, the competition results will not be published until the ECCV 2020 workshop. As described in the test guidelines, for each algorithm, the number of submissions is limited to a maximum of three uploads total in the challenge. If you submit multiple entries, the best results will be used as your final results for the competition. After the workshop, we will publish the competition results on the leaderboard. You can also navigate there by clicking on your algorithm’s name in your account.
I you want to rename your algorithm, please contact us.


  1. How do we handle empty predictions i.e cases in which our model does not predict any bounding box for that image ? The evaluation code assumes at least one prediction made.


Leave a Reply

Your email address will not be published. Required fields are marked *