Datasets:
image imagewidth (px) 1.75k 1.75k |
|---|
Rendered turntable photos of Google Scanned Objects
Turntable photo sequences rendered from a few objects of
Google Scanned Objects, for trying photo-to-3D reconstruction
without a camera. The photos are synthetic: each scanned, textured mesh stands on a light disc in
front of a light backdrop and is seen from one ring of 72 cameras 20 degrees above it (5 degree
steps), through a real lens model (the 3DLF apiCAM PRO lens with its radial distortion), at
1749 x 1155 pixels, with one light that keeps its place relative to the camera, a cast shadow,
1 px Gaussian blur and mild noise. Rendered by crisp3ds-dense render of
Crisp 3D Studio, which uses them as downloadable
example objects.
| Object | Title | Source model | Photos | Size |
|---|---|---|---|---|
| rhino | Rhino figurine (Schleich African Black Rhino) | Schleich_African_Black_Rhino | 72 | 230 MB |
| cereal_box | Cereal box (Van's Honey Nut Crunch) | Vans_Cereal_Honey_Nut_Crunch_11_oz_box | 72 | 231 MB |
Layout: <object>/rgb/<object>_<n>_rgb.png in capture order and
calibration/render-3dlf-lens.json (the lens the photos were rendered with, in Crisp3DS's
crisp3ds_lens_calibration_v1 format; exact for these photos). manifest.json lists every file
with its size and SHA-256.
How they reconstruct
With Crisp3DS's default command (crisp3ds-dense run --photos <object>/rgb --calibration calibration/render-3dlf-lens.json --output RUN: threshold masks, its own turntable solver), all 72
photos register; scored against the scanned mesh (F1 at 0.5 / 1 / 2 % of its diagonal): rhino
0.867 / 0.959 / 0.990, cereal box 0.833 / 0.890 / 0.969. Details:
docs/OTHER-IMAGE-SETS.md.
Source and license
The objects are from Google Scanned Objects, Copyright 2020 Google LLC, licensed
CC BY 4.0 (as stated in each model's
model.config and metadata.pbtxt). Dataset paper: Laura Downs, Anthony Francis, Nate Koenig,
Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B. McHugh, Vincent Vanhoucke,
"Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items", ICRA 2022,
https://arxiv.org/abs/2204.11918. Models downloaded from Gazebo Fuel (owner GoogleResearch).
These photos are licensed CC BY 4.0 as well. Changes: the meshes and textures were rendered into photographs; no mesh or texture is redistributed here. Product names and brands visible on the objects belong to their owners; their appearance here implies no endorsement.
- Downloads last month
- 53