jnle/wildlife_conservation_camera_trap_dataset

A camera trap dataset created as a subset of the Wildlife Conservation Society (WCS) camera trap dataset, af featured on: https://lila.science/datasets/wcscameratraps.

The subset contains 14,000 images split into a test set of 2,000 images and a pool of 12,000 images of 6 species of animal, namely pecari tajacu, loxodonta africana, aepyceros melampus, equus quagga, madoqua guentheri and crax rubra that can be used to generate various different splits. The dataset also includes 3 pre-done splits. These splits were created using a k-fold cross-validation approach with a focus on class balance, ensuring that each fold maintains a representative distribution of classes. It creates training and validation splits while calculating and saving label distributions to preserve class balance across all folds.

Note that the labels are in YOLO format, meaning the label fiels can be read as: class index, x_center, y_center, width, height.