Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
image
imagewidth (px)
256
512

Sparse Metric Anchors — YCB benchmark

The 42-object benchmark behind the paper Sparse Metric Anchors for a Single-View 3D Generative Prior: The Output Frame Is the Bottleneck (ISIR, Sorbonne Université, 2026). Code and paper: github.com/635jack/sparse-metric-anchors — its colab/reproduce.ipynb recomputes every table of the paper from this dataset on a CPU runtime.

The paper asks what limits the injection of a few metric measurements — tactile contacts, one depth map — into a pretrained single-view 3D generative model, and finds that it is the generator's output frame, not the measurements. This dataset holds everything needed to reproduce its numbers without regenerating anything: the conditioning images, the reference meshes, every contact set under every condition, the reprojected depth views, and the results of every campaign.

Contents

objects/<name>/image.png    single conditioning image, Blender render, 42 objects
objects/<name>/mesh.obj     reference mesh, normalised (max extent 2.0, centred)
contacts/<name>/*.pt        contact sets: occluded / visible / all at 32, 128, 512 (misplaced camera)
contacts_render/<name>/     palpation8, occluded, visible, all at 128 — drawn from the rendering camera
contacts_render_vispatch/   palpation sites on the visible face, 128 — drawn from the rendering camera
                            points; palpation sites at 4, 8, 16, 32 (128 points);
                            palpationN_M.pt = N sites, M points
contacts/<name>/dh116_{4,8}.npz   one contact per site with outward normals
contacts/strategies.npz     grasp-simulator contacts, 5 objects x 3 strategies,
                            with normals and finger ray origins
depth_rig/<name>/reelle/    true depth at canonical views 3, 6, 10, 26 (PNG)
depth_rig/<name>/reprojetee/  the same views reprojected from view 3 alone
results/*.json              every campaign result cited in the paper
split.json                  the 15 / 5 / 2 split used only by the learned baseline

The results files

Each JSON is keyed by object, then by condition and seed (baseline_s0, palpation8_128_s1, …). Each entry carries f2 (F-Score at 2 % of the bounding-box diagonal), f1, chamfer, num_components, largest_component_ratio, is_surface, and recall split into recall_visible / recall_occlude (decided from a misplaced camera: see below, not reliable).

File What it is
full42_render_combined.json constraint anchoring, 42 objects × 3 seeds, frame estimated by silhouette + residual, palpation sites drawn from the rendering camera (Table I)
full42_render_oracle.json same anchors, frame given by the reference mesh, reusing the unanchored passes above — an upper bound, not a system
full42_render_oracle_visible.json, _occluded.json, _vispatch.json the other cells of the 2 × 2 anchor-set design: uniform points on the visible face (the depth-camera control), uniform points on the hidden face, palpation sites on the visible face (Table III)
full42_combined.json, full42_oracle.json the earlier campaign, sites drawn from the misplaced camera — kept as a replication (Table I, last row)
render_rerun_analysis.json, anchor_set_factorial.json, hidden_side_recall.json every number of Table I, Table III and the recall per side, with the per-case values
dm4_resultats.json one depth map reprojected into four canonical views, with and without contacts (Table II)
guidance_campaign.json 32 / 128 / 512 contacts, occluded / visible / all, 5 objects
sites_4.json, palpation_combined.json, sites_16.json, sites_32.json 4 / 8 / 16 / 32 sites at 128 contacts
dh116_regime.json 8 or 4 contacts, with normals, with placement noise (Table IV)
strategies.json three grasp strategies differing in how many contacts fall on the occluded face (Fig. 2b)
poisson_temoin.json Poisson reconstruction from the same oriented contacts, no generative model (Table III)
plancher.json, plancher_gel.json eight identical repetitions, frame re-estimated / frozen (Sec. VI-G)
scoring_noise.json the frozen-frame meshes re-scored four times each: how much of that floor is the scorer
pose_*.json ICP, silhouette and combined placement of the contacts, displacement from exact
visibility_render_camera.json visibility recomputed from the camera that rendered the images: silhouette check, hidden fractions, Fig. 2b

Provenance and licence

The reference meshes are the fixed YCB meshes published by the LL4MA Lab (ll4ma-lab/ycb-fixed-meshes, CC BY 4.0), themselves derived from the YCB Object and Model Set (Calli et al., 2015). The renders, contact sets and reprojected depth maps are derived from those meshes and are released under the same CC BY 4.0 licence. Please cite YCB and credit the LL4MA Lab when using the geometry.

The contact sets are simulated on the reference meshes and are therefore exact — filtered for reachability, occlusion and fingertip clearance, but noiseless except where noise was added on purpose. No real sensor produced them. The paper says so in its limitations, and so should anything built on this data.

The renders come from a single camera position, (2.2, −2.2, 1.8) looking at the origin in Blender's Z-up world, with a procedural checkerboard material on a uniform background. Blender's OBJ importer first turns each Y-up mesh as (x, y, z) → (x, −z, y), so in the frame of mesh.obj that camera sits along (2.2, 1.8, 2.2), at 2.0 / 1.3 times that distance (the renderer normalised objects to 1.3, the meshes here are normalised to 2.0). The campaigns placed it at (2.2, −2.2, 1.8) in the mesh frame, 68° off: the occluded / visible contact sets, the palpation sites, part_occultee in strategies.json and the recall_visible / recall_occlude columns all carry that error. visibility_render_camera.json recomputes the visibility results from the right camera, and the paper reports both.

Citation

@unpublished{anchors2026,
  title  = {Sparse Metric Anchors for a Single-View 3D Generative Prior:
            The Output Frame Is the Bottleneck},
  note   = {ISIR, Sorbonne Universit\'e. Manuscript.},
  year   = {2026}
}
@inproceedings{calli2015ycb,
  title     = {The YCB object and model set: Towards common benchmarks for manipulation research},
  author    = {Calli, Berk and Singh, Arjun and Walsman, Aaron and Srinivasa, Siddhartha and Abbeel, Pieter and Dollar, Aaron M.},
  booktitle = {Int. Conf. on Advanced Robotics (ICAR)},
  pages     = {510--517},
  year      = {2015}
}
Downloads last month
42