Blender Synthetic Objects Dataset

The Blender Synthetic Objects Dataset is a controlled, reproducible image dataset for studying how object geometry, furniture type, surface material, and viewpoint affect visual representations. It contains 2,020 Blender-rendered PNG images covering 10 geometric shapes, 10 furniture models, 10 material presets, and 10 azimuth views. The released dataset is available on Zenodo, and the generation code is maintained in the GitHub repository. Contents Dataset at a glance Motivation Dataset composition Download Directory structure File naming convention Generation pipeline Reproducing the dataset Potential applications Recommended evaluation protocols Limitations License Citation Dataset at a glance Property Value Geometric shape categories 10 Furniture categories 10 Material presets 10 Views per item 10 Angular interval 36° Geometric-object images 1,000 individual + 10 overview Furniture images 1,000 individual + 10 overview Total images 2,020 Released image resolution 1920 × 1080 Image format PNG Rendering engine Blender Cycles Zenodo release v1.0.0 (14 July 2026) License CC BY 4.0 The balanced portion of the dataset follows: 2 domains × 10 categories × 10 materials × 10 views = 2,000 images 2 overview scenes × 10 views = 20 images ------------ Total 2,020 images Motivation Real-world image collections often combine changes in geometry, material, illumination, background, viewpoint, and object scale. This makes it difficult to isolate which visual factor caused a model's prediction or embedding to change. This dataset provides a compact controlled benchmark in which: geometry or furniture identity can be varied while material and viewpoint remain fixed; material can be varied while object identity and viewpoint remain fixed; viewpoint can be varied in regular 36° increments; object scale, centering, camera framing, and the studio environment are generated consistently; labels can be recovered directly from file names without a separate annotation tool. Dataset composition Geometric shapes Label Description Sphere UV sphere with a smooth curved surface Cube Six-sided regular cube Cylinder Circular cylinder with flat end caps Torus Ring-shaped torus Tetrahedron Four-faced triangular solid Cone Circular base tapering to one point Icosphere Sphere approximated by triangular faces Octahedron Eight-faced regular polyhedron HexPrism Prism with hexagonal end faces SquarePyramid Square base tapering to one apex Furniture models Label Category DesignChair Chair OfficeChair Office chair CadeirEstf Chair Sofa Sofa DesignSofa Designer sofa HomeDesk Home desk SimpleDesk Desk Table Table Wardrobe4Door Four-door wardrobe Wardrobe Wardrobe Material presets The material labels represent controlled Principled BSDF parameter presets. They are not measured material scans and do not use photographic texture maps. Label Base color (RGBA) Roughness Metallic Transmission Intended appearance rough_oak (0.30, 0.15, 0.05, 1.0) 0.80 0.00 0.00 Rough warm-brown wood clear_glass (1.00, 1.00, 1.00, 1.0) 0.00 0.00 1.00 Clear smooth glass polished_steel (0.80, 0.80, 0.80, 1.0) 0.05 1.00 0.00 Polished steel rusted_iron (0.20, 0.05, 0.01, 1.0) 0.90 0.50 0.00 Dark, rough iron-like surface solid_gold (1.00, 0.70, 0.10, 1.0) 0.15 1.00 0.00 Polished gold-like metal veined_marble (0.90, 0.90, 0.90, 1.0) 0.20 0.00 0.00 Smooth light stone-like surface carbon_fiber (0.05, 0.05, 0.05, 1.0) 0.40 0.30 0.00 Dark composite-like surface blue_fabric (0.12, 0.22, 0.55, 1.0) 0.95 0.00 0.00 Matte blue fabric-like surface brown_leather (0.28, 0.15, 0.08, 1.0) 0.45 0.00 0.00 Brown leather-like surface glossy_plastic (0.85, 0.05, 0.05, 1.0) 0.12 0.00 0.00 Glossy red plastic Download Zenodo hosts the versioned dataset archives and provides persistent identifiers and checksums. File in Zenodo v1.0.0 Contents Size MD5 object.zip Geometric-object renders 1.9 GB 063A4A2D696703C0E0F126464628C332 furniture.zip Furniture renders 1.6 GB 6D217C9C2FD780764DF32A6E5A94C3B7 Download page: https://doi.org/10.5281/zenodo.21355399 After downloading, extract both archives into the same project or data directory. Directory structure Blender_Synthetic_Objects_Dataset/ ├── README.md ├── object.py # geometric-shape generator and renderer ├── furniture.py # furniture importer and renderer ├── blend/ # source furniture .blend files │ ├── DesignChair.blend │ ├── OfficeChair.blend │ └── ... ├── object/ # 1,010 geometric-object images └── furniture/ # 1,010 furniture images The image directories are generated automatically when the scripts are run. File naming convention Individual geometric objects obj_<material>_<shape>_angle_<view>.png Example: obj_clear_glass_Sphere_angle_3.png This file contains a Sphere with the clear_glass material at view index 3, corresponding to an azimuth rotation of 3 × 36° = 108° from the initial view. Individual furniture models fur_<material>_<furniture>_angle_<view>.png Example: fur_blue_fabric_OfficeChair_angle_7.png This file contains an OfficeChair with the blue_fabric material at view index 7. Overview scenes The current scripts and local archives use: object/00_global_scene_angle_<view>.png furniture/00_global_scene_angle_<view>.png The original Zenodo v1.0.0 archive preview may display these overview files with domain prefixes (obj_00_... and fur_00_...). Individual-image naming is unchanged. Parsing labels in Python from pathlib import Path MATERIALS = ( "rough_oak", "clear_glass", "polished_steel", "rusted_iron", "solid_gold", "veined_marble", "carbon_fiber", "blue_fabric", "brown_leather", "glossy_plastic", ) def parse_filename(path): stem = Path(path).stem domain, remainder = stem.split("_", 1) material = next(m for m in MATERIALS if remainder.startswith(f"{m}_")) category_and_view = remainder[len(material) + 1:] category, view = category_and_view.rsplit("_angle_", 1) return { "domain": {"obj": "object", "fur": "furniture"}[domain], "material": material, "category": category, "view_index": int(view), "azimuth_degrees": int(view) * 36, } Overview images should be excluded before applying this parser. Generation pipeline Both domains are rendered using a common controlled workflow: Reset Blender to an empty scene. Create geometric primitives or import furniture from the supplied .blend files. Remove detected furniture display bases and platforms. Apply one of the 10 material presets to every mesh in the item. Normalize the item using its full world-space bounding-box diagonal. Recenter the geometry, place it on the ground plane, and aim the camera at its true center. Fit the camera separately for each view with a fixed safety margin so the item remains centered and uncropped. Render 10 evenly spaced azimuth views using Cycles with 64 samples. Render 10 additional overview views for each domain. The controlled factors are object identity, material, and azimuth. Camera elevation, studio environment, ground plane, normalization rule, and sampling configuration remain fixed within each generated domain. Reproducing the dataset Requirements Blender with Cycles support; the current code has been tested with Blender/Python API 5.0.1. Python support provided by Blender, or a compatible standalone bpy installation. Approximately 4 GB of free space for the generated PNG files, plus temporary rendering overhead. An NVIDIA GPU with OptiX support for the scripts as currently configured. The GPU backend is set to OPTIX in both scripts. Users of CUDA, HIP, oneAPI, Metal, or CPU rendering must adjust the Cycles device configuration before running the full generation job. Run with Blender From the repository root: blender --background --python object.py blender --background --python furniture.py Run with standalone Python and bpy If a compatible bpy module is installed: python object.py python furniture.py Run the commands from the repository root because furniture.py resolves the blend/ directory relative to the current working directory. Main configuration parameters Parameter Default Purpose TEST_MODE False When enabled, renders one view per item and skips overview scenes NUM_ANGLES 10 Number of azimuth views SPACING 6.0 Grid spacing for overview scenes TARGET_SIZE 1.5 Target bounding-box diagonal used for normalization FRAME_MARGIN 1.18 Camera fitting safety margin Cycles samples 64 Samples per rendered image Potential applications material-recognition benchmarking; geometric-shape and furniture classification; multi-view representation learning; image retrieval and visual-similarity analysis; evaluation of vision embeddings and foundation models; controlled lighting and reflection studies; synthetic-to-real transfer experiments; testing sensitivity to viewpoint, geometry, and surface appearance. Recommended evaluation protocols The dataset does not prescribe an official train/validation/test split. To avoid leakage, choose splits according to the intended research question: Material recognition: split by object or furniture identity so the same geometry does not occur in both training and test sets. Shape or furniture recognition: split by material and/or viewpoint. Viewpoint robustness: train on a subset of view indices and evaluate on held-out angles. Cross-domain transfer: train on geometric objects and evaluate on furniture, or vice versa, where label semantics permit. Always report the split rule, random seed, excluded overview images, and whether multiple views of the same item can cross split boundaries. Limitations The dataset is synthetic and does not reproduce the full variability of real cameras, environments, manufacturing processes, or material wear. Material classes are simple BSDF presets rather than measured reflectance data or detailed texture maps. Names such as veined_marble and carbon_fiber describe intended appearance categories; explicit veins or woven texture patterns are not modeled. Only 10 discrete azimuth views and one camera-elevation configuration are provided. Labels are encoded in file names; no official tabular annotation or benchmark split is included in v1.0.0. Furniture meshes come from third-party 3D-model sources and may contain differences in topology and geometric detail. The released images are 1920 × 1080, while the scripts rely on Blender scene defaults unless resolution is explicitly configured by the user. Synthetic-to-real conclusions should be validated on independent real-world data. The dataset contains no people, personal information, or sensitive attributes. License The released dataset is licensed under the Creative Commons Attribution 4.0 International License, as recorded on Zenodo. You may share and adapt the dataset with appropriate attribution. The original third-party furniture models remain subject to the terms of their respective source pages. Consult those sources before redistributing the source .blend assets separately. Citation If you use this dataset, please cite the version used. APA Dong, Y. (2026). Blender Synthetic Objects Dataset (Version 1.0.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.21355399 BibTeX @dataset{dong_2026_blender_synthetic_objects, author = {Dong, Yuhang}, title = {Blender Synthetic Objects Dataset}, year = {2026}, version = {1.0.0}, publisher = {Zenodo}, doi = {10.5281/zenodo.21355399}, url = {https://doi.org/10.5281/zenodo.21355399} } Furniture model sources antonio36. (2019, November 20). HomeDesk [3D model]. Free3D. https://free3d.com/3d-model/home-desk-235534.html berkayc. (2019, November 20). Sofa [3D model]. Free3D. https://free3d.com/3d-model/sofa-801691.html deiamasc. (2021, April 15). Cadeira [3D model]. Free3D. https://free3d.com/3d-model/cadeira-378801.html denakrom. (2020, February 19). SimpleDesk [3D model]. Free3D. https://free3d.com/3d-model/simple-desk-258372.html darpor, a. (2020, January 3). OfficeChair [3D model]. Free3D. https://free3d.com/3d-model/office-chair-480392.html hermes25. (2019, November 20). DesignChair [3D model]. Free3D. https://free3d.com/3d-model/design-chair-38271.html hermes25. (2019, November 20). DesignSofa [3D model]. Free3D. https://free3d.com/3d-model/design-sofa-85845.html kpawlowicz. (2019, November 20). SimpleWardrobe [3D model]. Free3D. https://free3d.com/3d-model/simple-wardrobe-153178.html mitsui. (2020, October 14). Table [3D model]. Free3D. https://free3d.com/3d-model/table-747735.html vinitk. (2019, November 20). Wardrobe4Doors [3D model]. Free3D. https://free3d.com/3d-model/wardrobe-4-doors-755350.html Repository Source code: https://github.com/SYadda/Blender_Synthetic_Objects_Dataset Dataset: https://doi.org/10.5281/zenodo.21355399