For each image of the training (1,407 images) and validation set (772 images), we provide the following annotations in the corresponding sub-folders:

• semantics: Pixel-wise semantic masks, where label ids correspond to: background (0), crop (1), weed(2), partial-crop (3), partial-crop (4). Partial crops and weeds have less than 50% visible pixels.
• plant_instances: Pixel-wise instance masks for crops and weed, where ids > 0 correspond to distinct instances.
• leaf_instances: Pixel-wise instance masks for leaves, where ids > 0 correspond to distinct instances.
• plant_visibility: Pixel-wise visibility in range [0,255], where 255 means fully visible.
• leaf_visibility: Pixel-wise visibility in range [0,255], where 255 means fully visible.

We store all annotations as 16-bit png files as these provide a decent lossless compression and can be easily read with off-the-shelf pillow or OpenCV, minimizing the dependencies.

All images and corresponding annotations have an image size of of size 1024 by 1024 pixels. The size was chosen such that even in later growth stages multiple plants are completely inside the image.

We additionally provide the acquisition data, e.g., 05-15, 05-26, or 06-05, at the beginning of every filename, which allows to separate the data based on the date of data acquisition.

Please, see our devkit providing a Pytorch dataloader that is ready to use with the dataset, but also the baselines implementations or instructions to reproduce our experiments in a separate repository. See our Code page for more information.