GOOSE-OCC: Semantic Occupancy and Observation-State Annotations for Monocular Off-Road Perception

It contains 3,367 pose-valid target frames, including 2,877 training frames and 490 validation frames, with a sequence-disjoint route-level split.Each target uses the front-facing CAM_WINDSHIELD_VIS camera and is associated with RGB imagery, 2D and 3D semantic annotations, LiDAR data, sensor calibration, vehicle pose, Metric3D pseudo-depth, and sparse semantic occupancy labels.

The occupancy grid covers 50 m x 32 m x 8 m at an isotropic resolution of 0.25 m, resulting in a grid of 200 x 128 x 32 voxels. The taxonomy contains seven occupied semantic classes plus free space.The release also provides causal temporal indices, history RGB images, SAM3 semantic pseudo-labels with confidence maps, history Metric3D pseudo-depth, and observation-state masks. The masks distinguish known free space, visible occupied voxels, occluded observed voxels, and unknown space, enabling separate evaluation of visible perception and occluded completion with valid free-space negatives.Under the accompanying weak-supervision protocol, occupancy ground truth and LiDAR-derived observation masks are used only for validation and evaluation, not as training losses.