ACFR Orchard Fruit Dataset

Fruit Dataset

The dataset - acfr-multifruit-2016 - contains images and annotations for different fruits, collected at different farms across Australia.
Download Link: acfr-multifruit-2016

The dataset was gathered by the agriculture team at the Australian Centre for Field Robotics, The University of Sydney, Australia. Further information about the group can be found at sydney.edu.au/acfr/agriculture

Data summary

Fruit	Image Info	Annotation Info	Notes
Apples	1120 Images PNG Image Data 308 x 202, 8-bit/color RGB Sensor: PointGrey Ladybug3	Circle Annotation (x,y,radius) Pixel-wise annotation	Collected at Warburton, Australia Apple varieties: Pink Lady, Kanzi
Mangoes	1964 Images PNG Image Data 500 x 500, 16-bit/color RGB Sensor: Prosilica GT3300c Strobes Used	Rectangle Annotation (x,y,dx,dy)	Collected at Bundaberg, Australia Mango varieties: Calypso
Almonds	620 Images PNG Image Data 308 x 202, 8-bit/color RGB Sensor: Canon EOS60D	Rectangle Annotation (x,y,dx,dy)	Collected at Mildura, Australia Almond variety: Nonpareil

Dataset Structure:

acfr-fruit-dataset
├── almonds
│   ├── annotations
│   ├── images
│   ├── labelmap.json
│   └── sets
│       ├── all.txt
│       ├── test.txt
│       ├── train.txt
│       ├── train_val.txt
│       └── val.txt
├── apples
│   ├── annotations
│   ├── images
│   ├── labelmap.json
│   ├── segmentations
│   └── sets
│       ├── all.txt
│       ├── test.txt
│       ├── train.txt
│       ├── train_val.txt
│       └── val.txt
├── mangoes
│   ├── annotations
│   ├── images
│   ├── labelmap.json
│   └── sets
│       ├── all.txt
│       ├── test.txt
│       ├── train.txt
│       ├── train_val.txt
│       └── val.txt
└── readme.txt

Each fruit contains an images, annotations and sets folder, along with a labelmap.json file for mapping labels to id.
The images contained within are random crops extracted from data otherwise densely collected over the farms. The image names are typically appended with the ij co-ordinate or image sections from which the particular crop was extracted from. Refer to [1] for further information on data gathering.
The Annotations folder contains corresponding .csv files with the same names as the images, with fruit annotation information per line.
The sets folder contains the dataset splits (list of image names) as used in [1] for training, testing and validation.
For apples, the segemntation folder contains png files with pixel-wise annotations of apples.

Results

The fruit detection F1-scores as computed in [1] are:

Fruit	F1-score
Apple	0.904
Mango	0.908
Almond	0.775

Refer to [1] for specifics on evaluation procedure.

[1] Bargoti, S., & Underwood, J. (2016). Deep Fruit Detection in Orchards. arXiv preprint arXiv:1610.03677. [Submitted to ICRA (2017)]

Annotation Info:

All object annotations were captured using PychetLabeller, which can also be used to quickly visualise the annotations (see code help page).

Both Almond and Mango datasets are annotated with rectangles, whereas the apples, being circular in nature, are annotated using Circles. The apple dataset contains additional pixel-level annotations.

To run the labeller:

python -m pychetlabeller apples/images/ apples/annotations/ --tool circle --labelmap apples/labelmap.json
python -m pychetlabeller mangoes/images/ mangoes/annotations/ --tool rectangle --labelmap mangoes/labelmap.json
python -m pychetlabeller almonds/images/ almonds/annotations/ --tool rectangle --labelmap almonds/labelmap.json

Citation

The dataset presented here has been used in multiple studies conducted at ACFR.
When using this dataset in your research, please cite the following papers.

For fruit level annotations:
@article{bargoti2016deep,
  title={Deep Fruit Detection in Orchards},
  author={Bargoti, Suchet and Underwood, James},
  journal={arXiv preprint arXiv:1610.03677},
  year={2016}
}

For pixel level segmentations (apple only):
@article{Bargoti2016,
  author={Bargoti, Suchet and Underwood, James},
  journal={To Appear in Journal of Field Robotics},
  title={Image Segmentation for Fruit Detection and Yield Estimation in Apple Orchards},
  year={2016}
}

Other Citations

Collection of all papers that have used part/all of the above data.

Fruit Annotations:

Bargoti, S., & Underwood, J. (2016). Deep Fruit Detection in Orchards. arXiv preprint arXiv:1610.03677. [Submitted to ICRA (2017)]
Stein, M., Bargoti, S., & Underwood, J. (2016). Image Based Mango Fruit Detection, Localisation and Yield Estimation Using Multiple View Geometry. Sensors, 16(11), 1915.

Pixel-level annotations (apples only):

Bargoti, S., & Underwood, J. (2016). Image Segmentation for Fruit Detection and Yield Estimation in Apple Orchards. arXiv preprint arXiv:1610.08120. [Accepted in Journal of Field Robotics, avaiting publication]
Bargoti, S., & Underwood, J. (2016). Image classification with orchard metadata. In Proceedings - IEEE International Conference on Robotics and Automation (Vol. 2016–June, pp. 5164–5170). http://doi.org/10.1109/ICRA.2016.7487721
Bargoti, S., & Underwood, J. (2015). Utilising Metadata to Aid Image Classification in Orchards. In IEEE International Conference on Intelligent Robots and Systems (IROS), Workshop on Alternative Sensing for Robot Perception (WASRoP).
Hung, C., Underwood, J., Nieto, J., & Sukkarieh, S. (2015). A Feature Learning Based Approach for Automated Fruit Yield Estimation. In Field and Service Robotics (FSR) (pp. 485–498). Springer. http://doi.org/10.1007/978-3-319-07488-7_33

Dataset Examples

Apple annotations have been converted to square bounding-boxes for visualisation.

Platform

The apple and mango dataset were collected using Shrimp - an unmanned ground vehicle, built at the Australian Centre for Field Robotics, The University of Sydney.
Figure below shows the vechicle operating in the apple and mango orchards.

Contact

For any questions, feel free to contact:

James Underwood: james.underwood@sydney.edu.au
Suchet Bargoti: suchet.bargoti@sydney.edu.au