Monastery Apple Dataset

To facilitate semi-supervised apple detection, we propose the Monastery Apple Dataset (MAD). The data was acquired in collaboration with a monastery in Bad Oldesloe, Germany. We collected video data from 16 trees of the monastery’s apple orchard with a DJI-Mini 3 Pro drone that has a resolution of 4k (3840 × 2160). The data collection was conducted over the course of one month in September 2022 and during different lighting conditions to ensure diversity. To generate the training, validation, and test splits, we first assigned the videos of 6 (training), 2 (validation), and 4 (test) trees to the respective splits. From the videos, we manually selected images for annotation to maximize data diversity. The training, validation, and test splits consist of 66, 12, and 27 frames with 10,089, 1,288, and 3,290 annotated instances as shown in the following table.

Split	Images	Labeled Instances
train	66	10,089
val	12	1,288
test	27	3,290
unlabeled	4,440	n/a

The apples were manually annotated with bounding boxes. An example image with ground truth from our dataset is shown below. Moreover, we automatically assigned three properties to each annotated apple representing the relative size, the level of occlusion, and the lighting conditions (see below). This enables us, to evaluate the influence of such conditions on our system. The 4,440 unlabeled images in the training split are composed of frames from the 6 training trees that were not annotated and frames from the four remaining videos/trees that were not assigned to the other splits.

If you use this dataset, please cite our paper:

@inproceedings{JohansonEtAlWACV2024,
title = {{S³AD}: Semi-supervised Small Apple Detection in Orchard Environments},
author={Johanson, Robert and Wilms, Christian and Johannsen, Ole and Frintrop, Simone},
booktitle = {Winter Conference on Applications of Computer Vision (WACV)},
year = {2024}
}