This dataset contains 34666 RGB-images taken from different angles and distances of weeds common in Manitoba. The imaged species common name, scientific name, and number of their images are:

Furthermore, this dataset contains a trained ResNet50 convolutional neural network model. It is trained to distinguish between monocots and dicots. A small collection of test datasets is included that can be used to measure the generalization capabilities of trained models.

The single-plant dataset and all test-datasets are accompanied by a csv-file containing filenames with respective labels.

The data was collected by the EAGL-I system, an autonomous robotic platform that takes images and labels them.

The plants are being placed in a volume that can be traversed by a camera on a robotic arm. Using the knowledge of the plant's position, the camera's position, and the label of the plant, the position of the plant in the image (in terms of pixels) can be calculated and thus a labeled bounding box created. This bounding box is cropped out to generate the single-plant images.

The above dataset is the "minimal" dataset used to train the ML model included.
The original uncropped images and additional metadata on the images (such as camera-position and -orientation) are being planned to be made accessible. For details, please contact the creators listed above and/or follow updates on this dataset.

Citation:

Cite the following article, when using this dataset or parts of it:

@article{10.1371/journal.pone.0243923,
    doi = {10.1371/journal.pone.0243923},
    author = {Beck, Michael A. AND Liu, Chen-Yi AND Bidinosti, Christopher P. AND Henry, Christopher J. AND Godee, Cara M. AND Ajmani, Manisha},
    journal = {PLOS ONE},
    publisher = {Public Library of Science},
    title = {An embedded system for the automated generation of labeled plant images to enable machine learning applications in agriculture},
    year = {2020},
    month = {12},
    volume = {15},
    url = {https://doi.org/10.1371/journal.pone.0243923},
    pages = {1-23}
}

Bugs:

In some folders the first master-image was incorrectly captured (the camera did not pan into position). Thus, not showing the plants that are suppossed to be showing. This only affected the first image taken in that folder. All other images in that folders are still correctly labeled. The folders containing that bug are: 20200312, 20200316, 20200318, 20200401, 20200406, 20200407.