ClutterNBV v1.0: interactive next-best-view reconstruction in tabletop clutter

ClutterNBV is a simulation benchmark for next-best-view (NBV) planning for 3D reconstruction in cluttered tabletop scenes. Existing tabletop clutter datasets are static captures from a fixed set of viewpoints; ClutterNBV instead provides physically settled scenes plus an interactive GPU environment (in the accompanying code repository) in which a free camera can move anywhere in a workspace and receives freshly rendered observations at every step, together with ground truth for the observable surface of every target object and of the whole table.

Two tasks share the same scenes:

• Global: reconstruct every object on the table (the setting of GenNBV-style NBV policies). Metric: coverage of the whole-table observable surface.
• Specified: reconstruct one target object, given by its segmentation mask in the observations. Metric: coverage of the target's observable surface.

Both use a budget of 10 views (the start view plus 9 actions) and fixed episode manifests (scene, target, start pose), so every method is evaluated on exactly the same episodes.

Preview

360° orbits of four test scenes (unseen objects), rendered by the benchmark environment: RGB, depth, per-object instance mask and the specified target object.

ring	ring_cover
packed	pile

Contents

Split	Objects	Scenes	Targets	Visibility matrices	Use
train	759 Google Scanned Objects	1,998	5,674	no	training (sizes S / M / L, protocol: M)
val	the training objects, new layouts	153	436	yes	model selection, fitting difficulty tiers
test	190 unseen objects	255	706	yes	final results

Objects are split at the object level. Scenes contain 7-16 objects (10.2 on average) in four layouts: ring (tall objects slid against the target), ring_cover (ring plus objects dropped onto the target), packed and pile; all scenes are settled with PyBullet and checked for stability and penetration.

Training-set sizes. splits/train_splits.json defines nested training sets S ⊂ M ⊂ L, stratified by layout and difficulty so that all three have the same mix (two-sample KS statistic of every difficulty statistic against L is at most 0.06). The official protocol trains on M, which fits a 40 GB GPU together with every baseline (measured on an A100 40 GB: scene pool 17.3 GiB + Hestia's 17.5 GiB peak) and a 24 GB GPU for the lighter GenNBV / FrontierNBV; S suits 16 GB GPUs and L (all training scenes) 80 GB GPUs. The pool costs about 24.9 MiB per scene plus 0.3 GiB.

Size	Scenes	Targets	Scene pool on the GPU
S	330	941	~8.3 GiB
M	700	1,975	~17.3 GiB
L	1,998	5,674	~48.8 GiB (extrapolated)

Ground truth

• Workspace (shared by the ground truth and the environment): camera within 1.0 m of the centre (the target centre for the specified task, the table-top centre for the global task), at most 70° elevation, at least 5 cm above the table and 5 cm from any geometry.
• Positions: 6 radii (0.2-1.0 m) x 700 Fibonacci directions, filtered by the workspace (about 2,400-2,800 per target or scene).
• Observable surface S_obs: surface samples seen with line of sight (range 0.15-1.5 m, incidence < 75°) from at least 5 positions, excluding points within 3 mm of another object or the table. Coverage is measured against S_obs, so contact faces never count as missed.
• Per target: 32,768 area-uniform samples of the object. Whole table: 16,384 area-uniform samples over all objects.

Episodes and difficulty tiers

File	Episodes
episodes/specified_train.json	22,694 (4 start poses per training target)
episodes/specified_val.json, specified_test.json	436, 706 (1 per target)
episodes/global_train.json	15,984 (8 start poses per training scene)
episodes/global_val.json, global_test.json	459, 765 (3 per scene)

Specified episodes use any target of the scene; the start camera looks at a random point of the clutter, sees 5-50% of the target's S_obs with at least 20 target pixels at 160 x 120, and its optical axis misses the target centre by at least 8°. Global episodes start from a valid pose in the table-centred workspace that sees 5-50% of the whole-table S_obs.

Every val/test episode carries a difficulty tier (easy / medium / hard) from random_cov, the expected coverage of the start view plus 9 uniformly random workspace views; thresholds are the tertiles on val (episodes/tiers_*.json) and are applied unchanged to test.

File format

• scenes/<split>/shard_XXXX.h5: one HDF5 group per scene id. attrs["recipe"] (JSON): table size, layout mode, objects with asset name and metric 6D pose, designated target. One subgroup target_KK per target with center, views (workspace positions), s_obs and contact (bit-packed), vis_count, the stored look-at start view and greedy oracle, attrs["stats"] (difficulty statistics), and on val/test vis, the bit-packed position x point line-of-sight matrix.
• scenes/<split>/scene_gt_XXXX.h5: whole-table ground truth per scene id: pts, nrm, owner (object index per point), s_obs, vis_count, views, and on val/test vis.
• assets/render_assets.h5: render meshes and grayscale textures per object; assets/render_textures_rgb.h5: the same textures in colour (optional RGB rendering); assets/gso_surfaces.tar: per-object surface samples (gso/<object>/surface.npz, extract next to assets_index.json).

Usage

The environment, baselines (GenNBV, FrontierNBV, Hestia) and evaluation scripts are in the accompanying code repository (link added at release). After downloading, set CNB_DATA to the dataset directory and extract assets/gso_surfaces.tar inside assets/.

License and attribution

The scene annotations, ground truth and manifests are released under CC BY 4.0. The object meshes and textures are derived from Google Scanned Objects (CC BY 4.0): L. Downs et al., "Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items", ICRA 2022.

Citation

To be added with the paper.