Dataset: Label-efficient phylloxera gall segmentation via 3D bottom-to-top leaf registration

This data repository contains relevant data created for the publication "Label-efficient phylloxera gall segmentation via 3D bottom-to-top leaf registration". For more details on the publication, submission status and code, see: https://github.com/eric-stumpe/Phyllox-Reg . The data repository contains the following directories: dataset - contains all recorded multimodal camera data of grapevine leaves infested by phylloxera, which were captured from the top and bottom sides. In addition, masks used for training DL-based segmentation are also part of this directory. hs_np: contains all recorded hyperspectral images (SpecimIQ) in .npy format hs: contains the same hyperspectral images as RGB representation in .png format rgb: contains RGB images recorded with an Orbbec Femto Bolt camera in .png format infrared: contains infrared images from the Orbbec Femto Bolt camera in .npy format depth: contains depth maps from the Orbbec Femto Bolt camera in .npy format thermal: contains thermal images recorded with a VarioCAM HDx Head 620 S thermal camera in .npy format (not used in publication) hs_mask_annot_bot: contains a binary image mask with hand-labeled phylloxera galls of bottom side leaf images hs_mask_annot_top: contains a binary image mask with hand-labeled phylloxera galls of top side leaf images hs_mask_transfer_direct: contains gall segment mask that have been transferred from bottom to top side with the developed registration algorithm hs_mask_transfer_dilated: contains gall segmentation masks that have been refined via dilation from hs_mask_transfer_direct hs_mask_transfer_peak: contains gall segment mask that have been refined via thresholding and pseudo labels from hs_mask_transfer_direct hs_pseudo_top: contains the raw top-side pseudo labels predicted by a Patch-CNN model that has been trained on bottom-side data (.npy) hs_pseudo_top_binary: contains hs_pseudo_top, but binarized at a value of 0.5 calibration_params - contains intrinsic and extrinsic camera parameters for each camera (hs: Specim IQ, rgbd: Orbbec Femto Bolt, thermal: VarioCAM HDx Head 620 S). Was calculated using checker board calibration. pretrained_weights - contains model weights (.pth - PyTorch) of three different models (MLP, Patch-CNN, UNET) that have been trained on bottom-side data.These models are then used for transfer learning.