Datasets from "A general method for bootstrapping dense 3D segmentations from sparse 2D annotations"

Image volumes used by the bootstrapper_nm_capsule (https://github.com/ucsdmanorlab/bootstrapper_nm_capsule), a reproduction of "A general method for bootstrapping dense 3D segmentations from sparse 2D annotations." The method turns a sparse 2D annotation budget into a dense 3D instance segmentation (pseudo ground-truth), then shows that a 3D network bootstrapped from that pseudo-GT approaches one trained on dense ground truth. This record is a single archive, data.tar (~3.9 GB). It unpacks to data/<dataset>/, one directory per dataset, each holding zarr (https://zarr.dev) arrays: - Core datasets (cremi_a, cremi_b, cremi_c, epi, fib, harris15) have volume_1.zarr/{raw,labels,sparse_labels} and a held-out volume_2.zarr/{raw,labels}.- Single-volume datasets (liconn, mitoem, cremi_clefts, fluo, prism) have only volume_1.zarr, with an added sparse_labels_mask. Modalities and voxel sizes (z, y, x in nm): cremi_a/b/c — EM Drosophila neuropil (40, 8, 8); epi — plant epithelium (235, 75, 75); fib — FIB-SEM, isotropic (8, 8, 8); harris15 — EM hippocampal neuropil (50, 8, 8); liconn — expansion-microscopy LM (24, 18, 18); mitoem — EM mitochondria (30, 8, 8); cremi_clefts — EM synaptic clefts (40, 4, 4); fluo — fluorescence, 2D+time (1, 1, 1); prism — PRISM expansion LM, 18-channel (400, 168, 168). Usage: download data.tar and run tar -xf data.tar (or use ./download_data.sh in the repo, which fetches this record and unpacks it into data/). Code, environment, and full run instructions are in the GitHub repository above. These are publicly available datasets, consolidated and reformatted to zarr for this reproduction. Please cite the original dataset sources as well as this archive.