This is a repository that contains links to the KOLOMVERSE dataset and codes to train and reproduce results of the paper titled "KOLOMVERSE: An open large-scale dataset for object detection in the maritime universe".
KOLOMVERSE is a large-scale object detection dataset in the maritime domain. It has a total of 186,419 4K resolution images categorized into 5 different classes namely ship (393,936 instances), buoy (34,080 instances), fishnet buoy (95,815 instances), lighthouse (60,362 instances) and wind farm (147,846 instances). The dataset is collected from 21 territotial waters of South Korea and is split in to train (49,175 images), validation (18,643 images) and test (18,601 images) sets.
If you wish to download the dataset along with the codes and scripts, please fill this Google request form. Once accepted, we shall mail you the link to download our dataset.
Distribution of train-test-validation split in the 21 territories of KOLOMVERSE. 
The KOLOMVERSE images are categorized into 5 different
classes namely hip (393,936 instances), buoy (34,080 instances), fishnet buoy (95,815 instances), lighthouse (60,362 instances) and wind farm (147,846 instances). 
KOLOMVERSE has variations in illumination, viewpoint, occlusion, background, scale and proportion that makes object detectors trained with KOLOMVERSE more suitable for real time applications.
Image samples with illumination variation:
.png)
Image samples with occlusion variation:
Image samples with backgorund variation:
.png)
Image samples with viewpoint variation:
Image samples with proportion variation:
