In this github repository we release the novel WeedAI dataset and our pseudo-labels and train, validation, evaluation splits associated with the Crop and Weed dataset (CAW) [1].
Michael Halstead, Patrick Zimmer, and Chris McCool. A Cross-Domain Challenge with Panoptic Segmentation in Agriculture The International Journal of Robotics Research
The WeedAI dataset is available here for download. It is available in the CoCo format and an example dataloader is available in the libs directory of this repository.
To use this dataloader you will need the following requirements (see requirements.txt):
- numpy==1.19.5
- torch==1.9.1
- torchvision==0.10.1
- torchcontrib==0.0.2
- pytorch-lightning==1.2.10
- scikit-learn==0.24.2
- scikit-image==0.17.2
- tensorboardX
- opencv-python==4.5.3.56
- pycocotools==2.0.7
- pyyaml
An example of how to run the dataloader and extract samples using the example config file (configs/WeedAI_plant.yml) is shown in the python script example.py. You should note that in the yml file your _dataset.location should be the root location prior to the datasets directory.
| Your directory location | datasets | WeedAI
To evaluate on the CAW dataset for this work we needed to generate pseudo-labels for panoptic segmentation. These labels were generated using the stem location supplied in the dataset and the connected components algorithm. These labels along with our training/validation/evaluation splits can be downloaded from here. Please note, these pseudo-labels are noisy as stated in the paper and are only generated for the background/crop class.
[1] Steininger, Daniel and Trondl, Andreas and Croonen, Gerardus and Simon, Julia and Widhalm, Verena. The CropAndWeed Dataset: A Multi-Modal Learning Approach for Efficient Crop and Weed Manipulation, Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2023