The P2S2 segmentation dataset: annotated in-field multi-crop RGB images acquired under various conditions

Abstract : Images play a vital role in crop phenotyping. Pixel-wise classification (into vegetation/background) or semantic segmentation is a critical step in the computation of several canopy state variables. Current state of the art methodologies based on convolutional neural networks are trained on data acquired under controlled environments. These models are unable to generalize to real-world dataset and hence need to be fine-tuned using new labels. This motivated us to create the P2S2 segmentation dataset – a collection of multi-crop RGB images from different acquisition conditions. We present here the dataset and state of the art results.