1st Tiny Object Detection Challenge

Organized by yingling - Current server time: Aug. 7, 2020, 8:52 p.m. UTC

First phase

Final test
April 20, 2020, midnight UTC

End

Competition Ends
July 25, 2020, 11:59 p.m. UTC

Welcome to the 1st Tiny Object Detection Challenge !

We organize the first large-scale Tiny Object Detection (TOD) challenge, which is a competition track: tiny person detection. For this track, we will provide 1610 images with 72651 box-level annotations. We provide 18433 normal person boxes and 16909 dense boxes in training set. These image are collected from real-world scenarios based on UAVs. The persons in TinyPerson are quite tiny, and the aspect ratio of persons in TinyPerson has a large variance. Since the various poses and viewpoints of persons in TinyPerson, it brings more complex diversity of the persons. There are some annotation rules: in TinyPerson, there are only one class "person". There are four conditions where persons are labeled as “ignore”: 1) Crowds, which are hard to separate one by one when labeled with standard rectangles; 2) Ambiguous regions, which are hard to clearly distinguish whether there is one or more persons; 3) Reflections in Water, and 4) Some objects are hard to be recognized as human beings, we directly labeled them as “ignore” too. When evaluating on test set, we mainly follow pedestrian detection rules. The test set contains 13787 person boxes and 1989 ignore regions in 786 images.

The baseline and benchmark code are avilable on Github

Due to the extension of submission deadline of rlq-tod, we decide to extend deadline of the challenge nearly one month, 25 July 2020(11:59PM)!!! And no more extension!!!
It is strongly recommended to submit a paper: Long Paper or Abstract(Short) !!!

Evaluation Criteria

We evaluate both AP(average precision) and MR(miss rate), but only AP50_tiny will be use to score and rank.

For more detailed experimental comparisons, the size range is divided into 3 intervals: tiny[2, 20], small[20, 32] and all[2, inf]. And for tiny[2, 20], it is partitioned into 3 sub-intervals: tiny1[2, 8], tiny2[8, 12], tiny3[12, 20]. And the three IOU thresholds(0.25, 0.50, 0.75) are used for performance evaluation. For example, AP50_tiny means only evaluate on tiny(size in [2,20]) objects and IOU threshold set to 0.50.

In leadboard, AP50_tiny, AP25_tiny, AP75_tiny, AP50_tiny1, AP50_tiny2, AP50_tiny3 and MR50_tiny are show as results.(But only AP50_tiny are use to score and rank).

The evaluation code are modified from COCO evaluation API(maskrcnn benchamrk version) and Citypersons evaluation code, which are open source on Github 

Terms and Conditions

This page enumerated the terms and conditions of the competition.

Award

  • 1st-Place Award: 1200 USD + Certificate
  • 2nd-Place Award: 800 USD + Certificate
  • 3rd-Place Award: 500 USD + Certificate

Reference

Xuehui Yu, Yuqi Gong, Nan Jiang, Qixiang Ye, Zhenjun Han, Scale Match for Tiny Person Detection, the Winter Conference on Applications of Computer Vision, 2020.

Final test

Start: April 20, 2020, midnight

Competition Ends

July 25, 2020, 11:59 p.m.

You must be logged in to participate in competitions.

Sign In