Object Detection in Aerial Images (ODAI)

A Contest on ICPR'2018

News

  • 2018-2-7 The website for ODAI on ICPR'2018 is online. New

Description

Object detection in Earth Vision, also known as Earth Observation and Remote Sensing, refers to localizing objects of interest (e.g., vehicles and airplanes) on the earth’s surface and predicting their corresponding land-use categories. The task of object detection in aerial images is distinguished from the conventional object detection task in the following respects:

  • The scale variations of object instances in aerial images are considerably huge.
  • Many small object instances are densely distributed in aerial images, for example, the ships in a harbor and the vehicles in a parking lot.
  • Objects in aerial images often appear in arbitrary orientations.

This contest, organizing on ICPR'2018, features a new large-scale image database of object detection in aerial images, named DOTA with nearly 3000 large-size images (4000 × 4000), which contain 15 categories.

Through the dataset and the tasks, we aim to draw attention from the a wide range of communities and call for more future research and efforts on the problems of object dection in aerial images.

Timeline

  • Extra test images for ODAI-18 available March 15, 2018
  • Submission open March 15, 2018
  • Submission deadline April 15, 2018
  • Submission of contest report April 30, 2018

Tasks

We propose two tasks for this contest, namely object detection with horizontal bounding box and object detection with oriented bounding box. Task1 uses the initial annotation as ground truth, while Task2 uses the generated axis-aligned bounding boxes as ground truth. We recommond you to test your algorithms by way of Task1, although the results from task2 are also of great practical value.

For more details and submission format, refer to the Tasks Page on DOTA.

Dataset

The detection tasks of ODAI are based on the DOTA dataset. Specificly, the Train and Validation sets are the same as DOTA-v1. However, the images of Test set are partly from DOTA-v1, other test images are not available currently. For the use of the data, please cite our article and follow the usage license described in DOTA. The data of the DOTA are available at DOTA Dataset Page.

NOTE: Except the train/val set of DOTA-v1, extra data is also allowed to train your detector, but you must give a description in your submission.

Organizers

  • Gui-Song Xia Professor at LIESMARS, Wuhan University, China
  • Xiang Bai Professor of the School of Electronic Information and Communications, Huazhong University of Science and Technology, China
  • Serge Belongie Professor at Cornell Tech and the Department of Computer Science at Cornell University, United States
  • Jiebo Luo Professor of Computer Science, University of Rochester, United States
  • Mihai Datcu Scientist with the German Aerospace Center (DLR), Germany
  • Marcello Pelillo Professor of Computer Science, Ca’ Foscari University of Venice, Italy
  • Liangpei Zhang Professor at LIESMARS, Wuhan University, China
  • Fan Hu Postdoctoral researcher at the Electronic Information School, Wuhan University, China



AmazingCounters.com