wsbench: trained models
All trained models of What Labels Are Worth Buying? Budget-Matched Weak Supervision for Cell Segmentation (Khalid, Naseem, Dengel, Ahmed; ISBI 2027, submitted). Code: https://github.com/nabeelkhalid92/wsbench Labels and splits: https://huggingface.co/datasets/nabeelkhalid92/wsbench-labels
Every model is a Mask R-CNN R50-FPN (detectron2), trained with one label strategy at one annotation budget and one seed, and selected on the validation set. Together they cover every strategy, budget and seed of the paper, plus the ablations and baselines.
Layout
<dataset>/<method>/<setting>/seed<k>/
model_best.pth weights (val-selected checkpoint)
config.yaml full detectron2 config used for training
run_info.json dataset, method, label file, best iteration, val score, training time
val_history.json validation score at every evaluation
score_thr.json score threshold chosen on val (for PQ / F1)
index.json lists all models.
<method> |
Strategy in the paper |
|---|---|
mrcnn_full |
Mask: full masks (budget_mask_B<s>), or all training masks (mask: the "Full, 6.6 B1" reference) |
ours |
Box (budget_box_B<s>) and Mixed (budget_mixed_B<s>): SAM2.1 pseudo-labels with the verification score |
pointsup |
Box+1pt (budget_pts1_B<s>): one labelled point per box (PointSup-based Point2Mask) |
mrcnn_boxrect |
box used directly as the mask (floor) |
boxinst, discobox, boxsnake, boxteacher |
box-supervised baselines (their own architectures) |
B<s> is the annotation budget in seconds (B1 = time to box every training cell: yeast 7826 s, DSB2018 164605 s,
LIVECell 7130032 s). LIVECell settings with a cell-type suffix (_skbr3, _bv2, _shsy5y) are the single-cell-type
study. Other names: mixednocal = Mixed without calibration, *_selsingle / *_seltoppred / *_acc0_noweight =
pseudo-label selection ablations, pts1sam = Box+1pt with SAM2.1 pseudo-labels.
Use
from huggingface_hub import hf_hub_download
ckpt = hf_hub_download("nabeelkhalid92/wsbench-models", "yeast/ours/budget_box_B1960/seed0/model_best.pth")
cfg = hf_hub_download("nabeelkhalid92/wsbench-models", "yeast/ours/budget_box_B1960/seed0/config.yaml")
With the code repository: python wsbench/train/train_net.py --dataset yeast --method ours --setting budget_box_B1960 --seed 0 --train-json <label file> --predict-only after placing the folder under workspace/runs/.
Inputs follow the converted datasets of the code repository (yeast and DSB2018 images are 8-bit, see its converters).
Licence
Models trained on LIVECell (livecell/) are derived from LIVECell (CC BY-NC 4.0): non-commercial use only,
with attribution to Edlund et al., Nat. Methods 2021. Models trained on yeast in microstructures (yeast/,
MIT, Reich et al. 2023) and DSB2018 / BBBC038 (dsb2018/, CC0, Caicedo et al. 2019) are released under the
MIT licence. See LICENSE.md.
Citation
@inproceedings{khalid2027labels,
title = {What Labels Are Worth Buying? Budget-Matched Weak Supervision for Cell Segmentation},
author = {Khalid, Nabeel and Naseem, Khola and Dengel, Andreas and Ahmed, Sheraz},
booktitle = {Proc. IEEE International Symposium on Biomedical Imaging (ISBI)},
year = {2027},
note = {submitted}
}