lws0111/scenefun3d_mask3d

SceneFun3D — Mask3D Aux (minimal)

Auxiliary data for the Fun3DU pipeline running on remote machines (e.g. H200). Pair with lws0111/scenefun3d_val (raw dataset).

Layout

proposals/{val,train}/{visit}.npz        # Mask3D ScanNet200 proposals (per visit)
                                         #   pred_masks (N_points, K) bool
                                         #   pred_classes (K,) int64
                                         #   pred_classnames (K,) U17
                                         #   pred_scores (K,) float32
query_struct/{val,train}.json            # ollama llama3.1 parsed structure
                                         #   per-desc {root_object, part, ...}
e5_selection/{val,train}.json            # CLIP per-query-FPS (k1=15, k2=35) frame ids

Note: preprocessed npy intentionally NOT uploaded

compute_coarse_masks.py uses xyz_full = (pred_masks's N_points × 3). This was historically the first 3 columns of preprocessed/{visit}.npy. To save 51 GB:

preprocessed npy's row order is identical to {visit}_laser_scan.ply (see models/mask3d/preprocess_scenefun3d.py). So on the remote, just read the PLY via parser.get_laser_scan(visit_id) and np.asarray(laser.points).

compute_coarse_masks.py auto-falls-back to this path when --preprocessed_dir is omitted.

Sizes

• val: 30 visits. train: 200 visits (lists in data/scenefun3d/benchmark_file_lists/{val,train}_scenes.txt)
• proposals: ~5.3 GB total
• json: ~6 MB total