Download README.md from lipeize/graspnet-infinite-object-selection: direct link, hf CLI and curl.
- Browser
- Download file 11.8 kB
-
https://huggingface.co/datasets/lipeize/graspnet-infinite-object-selection/resolve/main/README.md
- Command line
-
hf download hf://datasets/lipeize/graspnet-infinite-object-selection/README.md
-
curl -L -H "Authorization: Bearer $HF_TOKEN" -o README.md https://huggingface.co/datasets/lipeize/graspnet-infinite-object-selection/resolve/main/README.md
license: other
pretty_name: GraspNet-Infinite object selection pages
tags:
- robotics
- grasping
- 3d
GraspNet-Infinite object selection pages
Review pages for choosing which objects stay in the GraspNet-Infinite pool. There are 4,897 pool objects in 10 HTML files, and three more files of new candidates that are not in the pool yet: 408 dragons and birds, and 690 scanned objects in two files.
| Files | Objects | Count |
|---|---|---|
batch_01_of_10.html to batch_08_of_10.html |
Objaverse and Google Scanned Objects | 3,759, in files of 510 |
batch_09_of_10.html and batch_10_of_10.html |
Meta Digital Twin Catalog | 1,138, in two files of 569 |
candidates_dragon_bird_01_of_01.html |
new candidates, not in the pool: dragons and birds from Objaverse | 408, in one file |
candidates_scans_very_sharp_01_of_01.html |
new candidates, not in the pool: scanned objects with very sharp texture | 300, in one file |
candidates_scans_sharp_01_of_01.html |
new candidates, not in the pool: scanned objects with sharp texture | 390, in one file |
table_materials_01_of_01.html |
not objects: table-top materials for the scene renders, on one test scene | 51 materials (47 candidates and the 4 in use), in one file |
selected_objects.html |
every object kept on 10 November 2026, as thumbnails by sub-category | 4,243, in one file |
test_topup.html |
every object of the dataset, for topping up the three test sets by hand | 4,068, in one file |
Each file is self-contained. All its images are embedded, so it opens with a double-click and needs no server and no network.
Update 2026-11-09: 319 Meta Digital Twin Catalog objects left the pool, because they were the same item in another colour or print. The files were regrouped at the same time, so the Meta objects now sit in their own two files.
New candidates: dragons and birds, 2026-11-10
candidates_dragon_bird_01_of_01.html holds 408 objects that passed quality control and are waiting for a choice. None of them is in the pool. Keep means the object should be added, reject means it should not.
- Where they come from: a keyword search for dragons and birds over the Cap3D captions and Sketchfab names of Objaverse 1.0. 3,344 objects that had never been examined went through quality control, and 356 passed. Another 52 are objects that earlier rounds kept and that sat outside the pool; they passed the same checks again.
- What they passed: a closed mesh in at most three bodies, no copy of a GraspNet benchmark object, a single graspable object that matches its label, and a size between 4 and 27 cm.
- Two requirements added for this search: no low-poly models, and real textures only. Flat-coloured and single-colour objects are out.
- Order: the 39 dragons first, then the 369 birds by sub-category and species.
- Not applied: the limit of 25 objects per sub-category. The page shows every candidate that passed.
The file works like the others. It saves its choices separately and exports them as selection_candidates_dragon_bird_01.csv. index_candidates_dragon_bird.csv lists its objects in order.
New candidates: scanned objects, 2026-11-10
Two files hold 690 scanned objects that passed quality control and are waiting for a choice. None of them is in the pool. Keep means the object should be added, reject means it should not.
| File | Texture detail at the object's assigned size | Objects |
|---|---|---|
candidates_scans_very_sharp_01_of_01.html |
at most 0.20 mm | 300 |
candidates_scans_sharp_01_of_01.html |
over 0.20 mm and at most 0.30 mm | 390 |
- Where they come from: the objects that Objaverse++ labels as scanned. 31,116 scans that had never been examined were downloaded, 13,536 had a texture sharp enough to go on, and 1,002 passed quality control. The 312 with detail coarser than 0.30 mm are on neither page.
- Texture detail: the size of the finest detail the texture resolves on the object, measured along its surface. Each object shows this number on its mesh line. Both files hold textures that are sharp down to the single texel with at most 6 % of the surface blurred; the difference is how much surface one texel covers, so the second file holds larger objects.
- What they passed: a closed mesh in at most three bodies, no copy of a GraspNet benchmark object, a single graspable object, a real texture, no low-poly model, and a size between 4 and 27 cm.
- Left out on purpose: rocks, specimens, fossils, seashells, bones and minerals.
- Order: by family, then sub-category, with the sharpest first inside each sub-category.
- Not applied: the limit of 25 objects per sub-category.
The files work like the others. Each saves its choices separately and exports them as selection_candidates_scans_very_sharp_01.csv and selection_candidates_scans_sharp_01.csv. index_candidates_scans_very_sharp.csv and index_candidates_scans_sharp.csv list their objects in order.
Table-top materials, 2026-11-10
table_materials_01_of_01.html is not about objects. The scene pipeline draws the table top from four photographed materials. This file shows 47 more candidates from Poly Haven and the four in use, marked IN PRODUCTION. Keep means the material should go into the table library, reject means it should not.
- Test scene: every image is the same production scene (train_9 scene_000012, 9 objects) rendered by the scene pipeline with only the table top changed. Camera, objects, lights, HDRI and floor are the same in every image. Top: an overview. Bottom left: a closer view. Bottom right: bare table, with the texture map itself in the corner.
- Shortlist: chosen by eye from the Poly Haven catalogue. Left out: damaged or dirty floors, outdoor ground, rusty or corrugated sheet, wall panels and strong relief.
- Scale: each texture tile is the material's real size, from 0.27 m for cloth to 4 m for concrete. The four materials in use are set to 1.0 m.
- Colour: the pipeline's lighting lifts mid-tones, so a mid-brown walnut renders pale. This scene's lighting is typical for the pipeline.
- Gaps: Poly Haven has no plain plastic or laminate and no clean bare metal. The closest here are a wood-look vinyl, black rubber tiles and a painted blue steel plate.
- Order: by kind, so like materials sit together.
| Kind | Candidates | In use |
|---|---|---|
| wood | 23 | 2 |
| stone | 6 | 1 |
| concrete and plaster | 5 | 1 |
| tiles | 1 | 0 |
| fabric | 9 | 0 |
| metal | 1 | 0 |
| vinyl and rubber | 2 | 0 |
The textures are CC0. The file saves its choices separately and exports them as selection_table_materials_01.csv. index_table_materials.csv lists the materials in order. In it, uid is the Poly Haven id and final_cm is the tile size in cm.
All selected objects, 2026-11-10
selected_objects.html shows every object that was kept on the pages above, as thumbnails on one scrollable page.
| Count | |
|---|---|
| Kept from the pool | 4,008 |
| New, kept on the three candidate pages | 232 |
| Left undecided on their page, marked | 3 |
| Sub-categories | 743 |
- Layout: one block per sub-category, grouped by family. Pool objects come first in a block, new ones after them.
- New objects carry a
newmark. Each is filed under a sub-category that already exists in the pool: 174 kept their label and 58 were moved to the closest existing one. The earlier label shows on hover and in the export. - Select: click a thumbnail. Shift-click selects a range inside a sub-category. Each sub-category has
select allandnone. - Export:
Export selected CSVwritesselected_objects_selected.csv.Export all CSVwrites every object with aselectedcolumn.Import CSVbrings an exported file back. - Filters:
selected only,new only, a name filter and a jump list. - Hover a thumbnail for a larger view, the description, size, source and uid.
The selection saves in the browser. Export it before clearing browser data. index_selected_objects.csv lists every object with its sub-category, family and origin.
Test set top-up, 2026-11-11
test_topup.html is for choosing new test objects by hand. The three test sets lost objects to the deletions of 10 November and are topped up to 100 each.
| Set | Now | Missing |
|---|---|---|
| Easy | 80 | 20 |
| Medium | 87 | 13 |
| Hard | 85 | 15 |
- Layout: every object of the dataset as a thumbnail, by sub-category under its family. The current test objects carry their set and cannot be changed.
- Pick: choose a set in the header (Easy, Medium, Hard), then click objects to add them to it. A second click takes an object out. Choosing another set and clicking moves it. The header counts each set against 100.
scenesmark: the object appears in rendered training scenes (1,900 candidates). Picking it means those scenes must leave training or be rendered again. 922 candidates appear in no training scene; 994 more are Meta objects, which have no scenes and another licence.- Filters: hide objects in training scenes, hide Meta objects, hide current test objects, test objects only, one set only, a name filter and a jump list.
- Export:
Export picks CSVwritestest_topup_picks.csvwithuidandset.Import CSVbrings such a file back. - Hover a thumbnail for a larger view, the description, size, source, training group and uid.
Picks save in the browser. Export them before clearing browser data. index_test_topup.csv lists every object with its sub-category, family, current test set, training group and whether it is in training scenes.
How to use
Download the files into one folder and open batch_01_of_10.html in a desktop browser. Chrome or Firefox are recommended. Most files are 100 to 115 MB.
| Key | Action |
|---|---|
| Up arrow | keep, then move on |
| Down arrow | reject, then move on |
| Left / right arrow | previous / next object, no decision |
| Backspace | clear the verdict |
| Space | next undecided object |
| E | export the CSV |
| Up twice quickly, file 1 only | keep and highlight the object |
Choices save in the browser as you go. Export the CSV before clearing browser data or changing machine. The counter at the top right shows how many choices have not been exported. Import CSV restores an exported file. Each file saves and exports separately, as selection_batch_NN.csv. In file 1, Export highlights downloads the highlighted objects as highlights_batch_01.csv.
Choices made in the earlier 11-file version carry over: each new file picks up the earlier choices for its own objects the first time it opens in the same browser. If that does not happen, export the CSV from the earlier file and use Import CSV in the new files. Import applies the rows that belong to the file it is used in.
What each image shows
- Top: the object at its assigned size between a 17.8 cm banana and a gripper whose fingers open 10 cm, on a 1 cm grid with bold lines every 5 cm. The scale is the same for every object.
- Bottom: two close-ups of the object alone.
- Orange note: the object is shown from its merged trimesh copy, because that is what the scenes render for it (93 objects).
Within each group, objects are ordered by family, then sub-category, then uid, so similar objects sit together. index.csv lists which file and position holds each uid.
Licences
The images are renders of third-party 3D models, each under its source's terms:
- Objaverse objects: the licence of each model (CC BY or CC0 in this pool).
- Google Scanned Objects: CC BY 4.0.
- Meta Digital Twin Catalog: CC BY-SA 4.0 with an added clause that the material may not be sold or incorporated into a product to be sold.