LOAM-D

LOAM-D v1.0: A Multi-Domain Agricultural Semantic Segmentation Dataset LOAM-D is a large-scale semantic segmentation dataset for autonomous agricultural machinery, covering two operating domains: paved farmyards and open arable fields. It provides 1,512 human-annotated images captured by seven roof-mounted wide-angle cameras, complemented by 10,329 SAM-3 autolabels for training. The dataset uses a 121-class ontology that extends the off-road GOOSE taxonomy with agriculture-specific categories such as tillage states, crop residues, farm equipment, and implements. All labels are provided as single-channel PNG masks (class indices 0–120, unlabeled 255) with GOOSE-compatible coarse-12 and traversability mappings included. Files in this record Dataset LOAM-D-v1.0.zip — Full dataset archive: images (JPEG, 1280×720), manual/hybrid/auto label masks (PNG, class indices 0–120), canonical train/val/test splits, and per-image metadata. See the enclosed README.md for the directory layout and training configurations. ontology.json — Class index, name, and colour definitions (121 classes). classes.txt — Plain-text class list. coarse12_mapping.yaml — Mapping from 121 fine classes to 12 GOOSE-level groups. traversability_mapping.yaml — Mapping from 121 fine classes to the 4-class traversability set. index.csv — Per-image inventory (domain, camera, recording, resolution, split membership). pack_stats.json — Archive-level statistics. Pretrained baselines Model_Checkpoints_README.md — Description of all checkpoints, training configurations, and reproduction instructions. pidnet_s_*.zip — PIDNet-S checkpoints (Config A, Config B, Coarse-12, human-only control). segformer_b4_*.zip — SegFormer-B4 checkpoints (Config A, Config B, Coarse-12, full-121, GOOSE-only transfer).