The CropAndWeed Dataset

This repository provides utility scripts for the CropAndWeed dataset, a large-scale dataset for Precision Agriculture, consisting of highly variable real-world images and multi-modal annotations for a rich set of crop and weed categories. A thorough description can be found in the corresponding paper and supplementary material published at WACV 2023.

Annotation Format

The annotations consist of multiple directories for each dataset variant in the following formats:

• bboxes contains csv-files for each image with object instances defined as: Left, Top, Right, Bottom, Label ID, Stem X, Stem Y
• labelIds contains semantic masks for each image
• params contains the following additional parameters for each image:
  • moisture: 0 (dry), 1 (medium) or 2 (wet)
  • soil: 0 (fine), 1 (medium) or 2 (coarse)
  • lighting: 0 (sunny) or 1 (diffuse)
  • separability: 0 (easy), 1 (medium), 2 (hard)

The corresponding label IDs for each datset variant are specified in datasets.py. The names of all image and annotation files are prefixed either with ave or vwg refering to the Application and Experimental Sets, respectively, as described in the paper. The following 4-digit numbers specify the recording session, while the last 4 digits are the image id.

Setup

Run setup.py to download and extract all dataset images and annotations. By default, the script creates mapped annotations for all pre-defined dataset variants.

Dataset Variants

To create a specific dataset variant, use map_dataset.py to map the label IDs of one label specification to another. For instance, the command map_dataset.py --dataset_target CropsOrWeed9 will create new directories for bboxes and labelIds with annotations mapped to the predefined variant CropsOrWeed9. Custom dataset variants can be added to datasets.py and used analogously.

Bounding boxes are created as two distinct sets for training and evaluation purposes with the latter including a fallback Vegetation class containing tiny instances as well as excluded and ambiguous species. Mapped semantic masks include an additional background class and are only stored if they contain any mapped species.

Visualization

Use visualize_annotations.py to create a combined visualization for bounding boxes, stem positions and semantic masks for each annotated image. The dataset variant to be visualized and a filter for included images can be specified as parameters. For instance the command visualize_annotations.py --dataset CropsOrWeed9 --filter vwg-0328 will create visualizations for all images of the CropsOrWeed9 variant containing the relevant classes and matching the recording session vwg-0328. Bounding boxes and masks use the defined colors for each label from datasets.py, the fallback Vegetation class is included in black color. The assigned colors and numbers of occurences for each class are provided below each image.

Licence

The CropAndWeed dataset is released to academic and non-academic entities for non-commercial purposes such as academic research, teaching, scientific publications or personal experimentation (LICENCE).

Citing

If you use the CropAndWeed dataset for your research, please use the following BibTeX entry:

@InProceedings{Steininger_2023_WACV,
    author    = {Steininger, Daniel and Trondl, Andreas and Croonen, Gerardus and Simon, Julia and Widhalm, Verena},
    title     = {The CropAndWeed Dataset: A Multi-Modal Learning Approach for Efficient Crop and Weed Manipulation},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2023},
    pages     = {3729-3738}
}