LUCYD-PISCO training-set sources v5: in-focus plankton object library and object-free PISCO background textures
Abstract
LUCYD-PISCO training-set sources (v5) Source data needed to regenerate the synthetic training set used to retrain the LUCYD deconvolution network for PISCO (Plankton In Situ Camera Observation) extended-depth-of-field (TAG-lens) images. Code: https://git.geomar.de/team-optics/plankton/underwater-lucyd (sim/gen_trainset_v5.py; development notes in docs/RETRAINING_LOG.md). The training set itself (1500 pairs of 2560x2560 PNGs, ~5 GB) is not archived because it is fully reproducible from these two inputs plus the generator: tar -xf objects_v3.tar LUCYD_BG_POOL=bg_pool LUCYD_BOOST=1.30 LUCYD_GEN_OUT=lucyd_trainset_v5 \ python sim/gen_trainset_v5.py 1500 Contents objects_v3.tar — sharp object library (29,477 grayscale PNGs) In-focus plankton/particle crops on a white (255) background, cut from PISCO frames recorded with the TAG lens switched off (Tagamplitude 0, i.e. conventional narrow depth of field, no EDOF blur). Dataset 20250717_PISCO-SCHARF, 5,639 frames, PISCO2. Extraction (sim/extract_tagoff_v3.py --focus 2.0): 439,989 candidate objects were scored with an edge-width focus metric measured on the untouched raw frame (focus_px = contrast / p90(|grad| on the object boundary)); the 29,477 with focus <= 2.0 px (6.7 %) were kept. Other filters: area 120–40,000 px, bbox <= 1200 px, contrast >= 60, mask dark fraction 0.40. Hollow/ring-shaped (transparent) objects are deliberately kept. Objects are not taxonomically labelled; the majority are marine snow / aggregates, with copepods, appendicularians, rhizaria, etc. present. bg_pool/ — 72 object-free background textures (2560x2560 grayscale PNGs) Real camera background in the domain the network runs on: background-corrected ("cleaned") frames as produced by the PISCO segmenter (cleaned = bitwise_not(absdiff(img, max over 12 neighbouring frames)), frames binned 5120 -> 2560). Each texture is the per-pixel temporal median of 15 consecutive cleaned frames, which removes all drifting objects and leaves the fixed-pattern / haze field (the generator re-adds per-image grain, sigma 2.8–3.8). 12 textures per source, file name _ _ .png: source cruise / profile notes M181 METEOR M181 (2022) PISCO2 M202 METEOR M202 (2024) PISCO2 MSM126 MARIA S. MERIAN MSM126 (2024) PISCO2 SO298 SONNE SO298 (2023) PISCO2 SO308 SONNE SO308 (2024) PISCO2 ATAIR ATAIR/BSH 2026-04, profile 20260422-2051 PISCO2; the validation profile 20260423-1100 is NOT used here Frames from PIScO_benchmark_dataset_v2 (1000 consecutive raw frames per cruise) and the ATAIR 2026 campaign. Textures were computed with sim/export_bg_pool.py (seed 0), identical to the pool built internally by gen_trainset_v5.py for the v5/v5d models. Generator summary (for context) For each training image: 25–55 objects drawn from the library, random rotation/scale (0.5–1.6, INTER_AREA when shrinking) and opacity (0.85–1.0), composited multiplicatively. GT = objects on pure white with contrast boost 1.30. Input = the same objects passed through the TAG-lens blur simulator (sim/edof_sim_gen_image.py, random depth position), composited onto a random background texture from bg_pool/, plus Gaussian grain. Checksums See sha256sums.txt.