Data sources

Where the ground truth comes from, what kind of ground truth it is, and what it costs you to use it. gt_mesh is the honest answer to "can I measure geometry with this?"

Procedural (exact GT, offline)

local · tier A · object + scene · GT: exact

Meshes are constructed analytically and rendered by the bundled numpy rasteriser, so mesh, depth, normals and poses are exact and mutually consistent. No SfM, no scan noise, no scale ambiguity. The reference tier for measuring a method's geometric ceiling.

licence: Generated locally -- no third-party licence
home: https://github.com/ (see docs/extending.md)

Objaverse 1.0

objaverse · tier A · object · GT: authored

800k+ annotated objects with varying mesh quality and topology. Use the LVIS subset for a category-balanced draw. Great breadth, mixed quality -- always sanity-check watertightness before trusting it as geometry GT.

licence: Per-object (CC0 / CC-BY / CC-BY-NC) -- check metadata
cite: Deitke et al., Objaverse, 2023
home: https://objaverse.allenai.org/

Google Scanned Objects (GSO)

gso · tier B · object · GT: scanned

1030 structured-light scans of household items (~13 GB). Real albedo and real scan noise on a clean mesh; single capture lighting, so pair it with synthetic relighting for the lighting axis.

licence: CC-BY 4.0
cite: Downs et al., Google Scanned Objects, 2022
home: https://research.google/blog/scanned-objects-by-google-research-a-dataset-of-3d-scanned-common-household-items/

Amazon Berkeley Objects (ABO)

abo · tier A · object · GT: authored

7,900+ artist-authored product models with 4K PBR textures, plus 398k catalogue photos. The strongest open source of LIGHTING VARIATION: every product is rendered under multiple HDRI studio environments, so the same geometry can be scored under very different illumination.

licence: CC BY-NC 4.0 (non-commercial)
cite: Collins et al., ABO, 2022
home: https://amazon-berkeley-objects.s3.amazonaws.com/index.html

Blender / NeRF synthetic

blender_synthetic · tier A · object · GT: exact

lego, chair, drums, ficus, hotdog, materials, mic, ship -- 100-200 exact poses in transforms.json plus the .blend files, from which the GT mesh is exported exactly. The de-facto reference object-level geometry benchmark.

licence: MIT / research (see release)
cite: Mildenhall et al., NeRF, 2020
home: https://jonbarron.info/mipnerf360/

ShapeNetCore v2.1

shapenet · tier A · object · GT: authored

51k models in 55 categories with consistent orientation, but no real images and a single shading setup. Use as a large geometry-only pool and render the lighting yourself (BlenderProc).

licence: Custom (registration required)
cite: Chang et al., ShapeNet, 2015
home: https://shapenet.org/

Replica

replica · tier C · scene · GT: scanned

18 photorealistic indoor scenes, each with a dense textured mesh (mesh.ply), HDR textures and semantics, at room and building scale. The best open scene-level GT mesh for indoor geometry evaluation.

licence: Research / non-commercial
cite: Straub et al., Replica, 2019
home: https://github.com/facebookresearch/replica-dataset

Tanks and Temples

tanks_and_temples · tier C · scene · GT: laser

Large real outdoor/indoor scenes with laser-scan ground truth supplied as dense point clouds. Registration to the GT frame is mandatory, so always report Sim3/ICP-aligned numbers.

licence: Research (manual download)
cite: Knapitsch et al., Tanks and Temples, 2017
home: https://www.tanksandtemples.org/download/

DTU MVS 2014

dtu · tier B · object + scene · GT: laser

124 robot-mounted tabletop scenes with STRUCTURED-LIGHT ground truth (point clouds), 49-64 views and calibrated lighting. The standard MVS geometry benchmark; its 22-scan protocol is directly reusable.

licence: Research
cite: Aanaes et al., DTU MVS, 2016
home: https://roboimagedata.compute.dtu.dk/?page_id=36

Mip-NeRF 360

mipnerf360 · tier C · scene · GT: none

Nine unbounded indoor/outdoor scenes. CAVEAT: no ground-truth mesh ships with it -- it was built for novel-view synthesis. It is registered here for appearance/robustness checks only. The COLMAP cloud it provides is reconstruction-derived and must NOT be used as geometry ground truth.

licence: CC BY 4.0
cite: Barron et al., Mip-NeRF 360, 2022
home: https://jonbarron.info/mipnerf360/

Which samples use which source

source	samples
`local`	obj_sphere_blob, obj_vase, obj_lego_brick, obj_capsule_pill, obj_torus_ring, obj_chair, obj_lamp, obj_studs_tile, scene_tabletop, scene_room_corner
`objaverse`	objaverse_lvis
`gso`	gso_household
`abo`	abo_studio
`blender_synthetic`	blender_hotdog, blender_lego, blender_ficus
`shapenet`	shapenet_chair
`replica`	replica_office
`tanks_and_temples`	tnt_family
`dtu`	dtu_scan1
`mipnerf360`	mipnerf360_bonsai