Dataset of Canada goose images with annotations of bounding boxes with object classes, suitable for testing object detection algorithms.
This is a dataset of 1,000 Canada goose images. They are all mugshots, with the head being the main feature of the image. Each image comes with an annotation XML in the PASCAL VOC format with a bounding box of the head and the object class goose-head.
Canada geese have a distinct and high-contrast facial pattern that is highly recognizable. The images in this dataset are all somewhat visually similar. The head in the image is either the left side or the right side, not front or back. The background is usually blurry. Only one goose head is present in each image. These properties makes the dataset rather simple for various computer vision tasks, and can serve as a starting point for computer vision experiments such as object recognition, localization, detection, etc.
All images are photographed by me.
The images directory contains 1,000 JPG images of the size 800 x 533. The annotations directory contains the XML annotations. These can be directly used in the darkflow implementation of object detection by placing them under darkflow/test/training.
If you are only using the images and not the annotations, I also implemented a load_data() API that has compatible usage as those in the Keras built-in datasets.
(x_train, y_train), (x_test, y_test) = GooseDataset.load_data(test_ratio=0.2,
grayscale=False,
cropped=False,
resize_shape=None,
save=False)
Returns:
Arguments:
To see an actual example, please check the Example Jupyter notebook.
Please do not use this dataset to derive any work that harms animals.