The Aerial Elephant Dataset: A new public benchmark for aerial object detection.

p, li { white-space: pre-wrap; } <pre>The dataset consists of 2101 images containing a total of 15 511 elephants. It is split into training and test subsets with 1649 images containing 12455 elephants in the training set and 452 images containing 3056 elephants in the test set. The resolution of the images varies between 2.4 cm/pixel and 13 cm/pixel, but the nominal resolution for each image is specified in the accompanying metadata, so it is a simple matter to resample images to a consistent GSD. Because acquired images often overlap, the same individuals may sometimes be seen in 2 or 3 consecutive images. Care has been taken with the train/test split to ensure that such clusters of related images are not split, thus maintaining independence of the training and test sets. </pre> <pre><br></pre> <pre>These images were acquired over the course of 8 separate campaigns in different environments.</pre><pre><br></pre>

本数据集共包含2101张图像，总计15511头大象。数据集划分为训练集与测试集两个子集：训练集包含1649张图像，对应12455头大象；测试集包含452张图像，对应3056头大象。图像的分辨率范围为2.4 cm/像素至13 cm/像素，但每张图像的标称分辨率均在配套元数据中注明，因此可便捷地将图像重采样至统一的地面采样距离（Ground Sample Distance）。由于采集的图像常存在重叠，同一头大象有时会出现在2或3张连续的图像中。在划分训练集与测试集的过程中已进行严谨处理，确保此类关联图像簇不会被拆分，以此保证训练集与测试集的独立性。这些图像是在8个不同环境下开展的独立采集任务中获取的。