Loading...
Loading...
Contributors:
Dataset URI:
10.5281/zenodo.23566224Synthetic rhizotron root dataset, Root LoRA weights and trained segmentation models==================================================================================== This record accompanies the paper "Diffusion-Based Synthetic Pretraining for RootSegmentation in Rhizotron Images". Code: see the code repository linked in the paper(during review: https://anonymous.4open.science/r/synthetic-rhizotron-roots-BB93/). FILES----- synthetic_dataset.zip 3,758 synthetic rhizotron images with their root segmentation masks (1024 x 1024 pixels, PNG). synthetic_dataset/ images/ 3,758 synthetic images labels/ 3,758 root masks Files follow the nnU-Net v2 naming convention: the image images/synthetic_0001_002_1_0000.png belongs to the label labels/synthetic_0001_002_1.png (the image has the extra channel suffix "_0000"). Images are single-channel greyscale. Labels have the values 0 (background) and 1 (root), so they look black in a normal image viewer. The folders can be used directly as imagesTr / labelsTr of an nnU-Net raw dataset. For every segmentation model in the paper, the images were split randomly into 80% training (3,006) and 20% validation (752) (see the code repository). The images were generated with Stable Diffusion 1.5 (v1-5-pruned-emaonly), the Root LoRA below and ControlNet v1.1 inpaint (control_v11p_sd15_inpaint), by inpainting roots into real soil-background patches, guided by composed root masks. The masks are therefore exact labels for the generated roots. All images in this dataset are synthetic: they were produced with generative AI and are not experimental observations. root_lora.safetensors Root LoRA (rank 128, alpha 64) for Stable Diffusion 1.5, trained with kohya_ss sd-scripts on rhizotron root patches. Load it in ComfyUI from models/loras/ together with the workflow in the code repository (04_synthetic_image_generation/root_inpainting_workflow_api.json). trained_models.zip Trained segmentation models for the three architectures and the three training strategies evaluated in the paper: trained_models/ nnunet/ Synthetic_only/ synthetic_only_checkpoint_final.pth, plans.json, dataset.json, dataset_fingerprint.json Real_only/ Real_only_checkpoint_final.pth, (same .json files) Finetuned/ finetuned_checkpoint_final.pth, (same .json files) yolo/ Synthetic_only.pt Real_only.pt Finetuned.pt mask2former/ Synthetic_Only.pth Real_only.pth Finetune.pth Strategies: Synthetic only trained on the synthetic dataset only Real only trained on the real training data only (100%) Fine-tuned pretrained on the synthetic dataset, then fine-tuned on the real training data (100%) To predict with an nnU-Net model, place its checkpoint in a fold_0 folder inside the model folder and point nnU-Net to it, e.g.: mkdir -p nnunet/Real_only/fold_0 mv nnunet/Real_only/Real_only_checkpoint_final.pth nnunet/Real_only/fold_0/ nnUNetv2_predict_from_modelfolder -i <images> -o <output> \ -m nnunet/Real_only -f 0 -chk Real_only_checkpoint_final.pth \ --save_probabilities Models: nnU-Net v2 (2D, fold 0; nnU-Net checkpoint files), YOLO26s-seg (Ultralytics weights) and Mask2Former with a ResNet-101 backbone (detectron2 weights; use with 05_segmentation/mask2former/configs/ maskformer2_roots.yaml from the code repository). Prediction and evaluation scripts are in 05_segmentation/ of the code repository. Predicted root probabilities were binarised at 0.01. REAL DATA--------- The real rhizotron images used to train the Root LoRA and the segmentationmodels, and used for evaluation, are subject to a confidentiality agreement withan industrial partner and are not included in this record. LICENSE------- Creative Commons Attribution 4.0 International (CC BY 4.0).Stable Diffusion 1.5 and ControlNet are subject to their own licences(CreativeML Open RAIL-M); the Root LoRA is intended for use with those modelsunder their terms.
Keyword(s):
MeSH Term(s):
Grants and Funding:
No Grants or Funding Data Available
Associated Publication(s):
No Associated Publication Data Available
Supplementary Data:
No Supplementary Data Available
Related Dataset(s):
No Related Dataset Data Available
The National Library of Medicine (NLM) provides information about biomedical datasets. Inclusion on an NLM Web page does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: NLM Web Policies (nih.gov)