Datasets:
3DReflecNet: A Large-Scale Dataset for 3D Reconstruction of Reflective, Transparent, and Low-Texture Objects
Online Browsing
We provide online browser for previewing a subset of 3DReflecNet, enjoy!
Dataset Summary
3DReflecNet is a large-scale synthetic multi-view dataset for novel view synthesis and 3D reconstruction of reflective, transparent, and low-texture objects. It provides Blender-rendered RGB images, masks, depth maps, normal maps, camera parameters, 3D model assets, and material/environment annotations.
The Hugging Face release uses a hybrid layout:
- WebDataset shards contain the frame payloads used for training.
- Metadata Parquet contains one row per frame for filtering, indexing, camera parameters, and flattened tags.
- Preview Parquet contains a small sampled subset with PNG preview bytes for the Dataset Viewer and Space demo.
- Model assets are uploaded as regular files and referenced by path from the metadata.
Release Structure
data/
├── metadata/
│ ├── train.parquet
│ ├── test.parquet
│ └── parts/
│ ├── train/
│ │ └── {release_name}.parquet
│ └── test/
│ └── {release_name}.parquet
├── preview/
│ ├── preview.parquet
│ └── parts/
│ └── {release_name}.parquet
├── shards/
│ ├── train/
│ │ └── {release_name}/
│ │ ├── train-000000.tar
│ │ └── ...
│ └── test/
│ └── {release_name}/
│ ├── test-000000.tar
│ └── ...
└── models/
└── {release_name}/
├── {number}_file/
│ ├── {number}_file.obj
│ ├── {number}_file.ply
│ └── {number}_file.glb # preview instances only
└── {number}_blend/
└── {number}_blend.blend
OBJ, PLY, and BLEND assets are uploaded for each released model. GLB files are generated only for preview instances for browser-based viewing in the Dataset Viewer or Space demo. For non-preview rows, glb_path may be an empty string.
Source Data Layout
The build scripts expect each source root to use this layout:
{DatasetRoot}/
├── {number}_file/
│ ├── {number}_file.obj
│ └── {number}_file.ply
├── {number}_blend/
│ └── {number}_blend.blend
└── {number}_{HDRI}_{Material}/
├── train/
│ └── frame_XXXXX.png
├── mask/
│ └── frame_XXXXX.png
├── depth/
│ └── frame_XXXXX.exr
├── normal/
│ └── frame_XXXXX.exr
├── test/
│ ├── rgb/
│ │ └── frame_XXXXX.png
│ ├── mask/
│ │ └── frame_XXXXX.png
│ ├── depth/
│ │ └── frame_XXXXX.exr
│ └── normal/
│ └── frame_XXXXX.exr
├── transforms_train.json
├── transforms_test.json
├── meta.json
└── tags.json
WebDataset Samples
Each WebDataset sample represents one rendered frame. Sample keys use:
{instance_id}_frame_{frame_id:05d}
Each sample contains:
| Member | Description |
|---|---|
{key}.rgb.png |
RGB render |
{key}.mask.png |
Object mask |
{key}.depth.exr |
32-bit floating point depth map |
{key}.normal.exr |
32-bit floating point normal map |
{key}.camera.json |
Per-frame camera data extracted from transforms JSON |
{key}.meta.json |
Scene metadata |
{key}.tags.json |
Material/environment annotations |
camera.json contains split, frame_id, source file_path, transform_matrix, and camera intrinsics.
Rendering and Generation Details
Scenes are rendered with Blender 4.5.5 LTS using the bundled BlenderNeRF-based rendering pipeline. The default renderer is Cycles with 64 samples, GPU denoising enabled, AgX view transform, and 1000 x 1000 output resolution.
Each released frame includes:
| Output | Format | Description |
|---|---|---|
| RGB | PNG | Blender-rendered color image. |
| Mask | 8-bit grayscale PNG | Object mask generated from Blender's Object Index pass. Dataset objects use pass index 1; helper camera/empty objects use pass index 0. |
| Depth | 32-bit grayscale OpenEXR | Raw Blender Depth/Z render pass. The EXR stores the scalar depth channel as V. Preview PNGs are normalized for visualization only. |
| Normal | 32-bit RGB OpenEXR | Blender Normal render pass stored in X, Y, and Z channels. Values are expected to lie in [-1, 1]. Preview PNGs map this range to [0, 255] for display only. |
PNG outputs use Blender native compression level 15. OpenEXR outputs use PIZ compression. The full training payload keeps depth and normal maps as EXR files; depth_preview and normal_preview in the preview parquet are display-only PNG conversions and should not be used as training targets.
The default generation configuration renders train views from the BlenderNeRF camera-on-sphere trajectory and test views from transforms_test.json. Metadata stores the corresponding camera-to-world transform matrix and intrinsics for each frame.
Metadata Parquet Fields
data/metadata/train.parquet and data/metadata/test.parquet contain one row per frame after partition metadata is merged. Per-archive metadata parts are stored under data/metadata/parts/{split}/{release_name}.parquet.
Important fields include:
| Field | Description |
|---|---|
release_name |
Archive/root partition name |
sample_id |
{split}/{sample_key} |
instance_id |
Scene instance identifier |
sample_key |
WebDataset sample key |
shard_path |
Tar shard containing the sample |
split |
train or test |
frame_id |
Frame index |
rgb_member, mask_member, depth_member, normal_member |
Member names inside the shard |
camera_member, meta_member, tags_member |
JSON member names inside the shard |
source_*_path |
Original source path for traceability |
transform_matrix |
4x4 camera-to-world matrix |
intrinsics |
Camera intrinsics struct |
blend_path, obj_path, ply_path, glb_path |
Model asset paths; glb_path is populated only for preview instances |
main_category, sub_category |
Object category labels |
model_name, material_name, env_name |
Scene descriptors |
hasGlass, isGenerated, transparent, near_light |
Boolean scene attributes |
glossiness, roughness, reflectivity, texture_scale, anisotropy, metallic_hint |
Flattened material tags |
indoor_outdoor, light_type, light_intensity, key_light_direction, shadow_hardness |
Flattened environment tags |
description |
Natural language scene description |
Preview Parquet
data/preview/preview.parquet is a merged small sampled subset used for browsing. Each build writes data/preview/parts/{release_name}.parquet; the current build rule is to randomly sample --preview_instances scene instances from each input archive or root using --preview_seed (default: 42). The Space demo loads only the merged preview parquet, not the full metadata parquet. Preview rows include the metadata fields above plus:
rgbmaskdepth_previewnormal_preview
These are PNG bytes intended for display only. Full training data should be read from WebDataset shards.
Helper Scripts
Optional helper scripts can be uploaded under scripts/ to download selected release partitions and extract WebDataset shards into local frame folders. These scripts are convenience utilities; the canonical dataset payload remains the metadata parquet files and WebDataset shards under data/.
Install helper script dependencies:
pip install -r scripts/requirements.txt
Download one train partition:
python scripts/download_3dreflecnet.py \
--output_dir 3dreflecnet \
--split train \
--release_name {release_name} \
--shard_index 0
Download only model assets for one release:
python scripts/download_3dreflecnet.py \
--output_dir 3dreflecnet \
--release_name {release_name} \
--models_only
Extract frames and generate display PNGs from depth/normal EXR files:
python scripts/extract_webdataset.py \
--dataset_dir 3dreflecnet \
--output_dir extracted \
--split train \
--release_name {release_name} \
--shard_index 0 \
--preview_png
Extract only RGB PNGs and the original 32-bit depth EXR maps:
python scripts/extract_webdataset.py \
--dataset_dir 3dreflecnet \
--output_dir extracted \
--split train \
--release_name {release_name} \
--shard_index 0 \
--modalities rgb depth
Supported modality names are rgb, mask, depth, normal, camera, meta, and tags. Selecting modalities reduces the extracted output size, but the complete selected tar shard must still be downloaded because all modalities are stored together in each WebDataset shard.
The extraction script keeps original depth/*.exr and normal/*.exr files by default. Use --png_only to write preview PNGs instead of EXR files for selected depth/normal modalities; other selected modalities are unchanged. These PNGs are visualization-friendly conversions and should not be treated as replacements for the original EXR training targets.
Camera Parameters
Camera extrinsics follow the NeRF / instant-ngp convention:
transform_matrix: 4x4 camera-to-world matrix- OpenGL convention: +X right, +Y up, -Z forward
- Intrinsics are copied from the corresponding
transforms_train.jsonortransforms_test.json
Intended Use
This dataset is intended for:
- Novel view synthesis for reflective/specular objects
- 3D Gaussian Splatting and NeRF benchmarking
- Depth estimation under challenging reflective surfaces
- Normal estimation for metallic and glossy materials
- Material-aware scene understanding
Citation
@inproceedings{liang20263dreflecnet,
title={3DReflecNet: A Large-Scale Dataset for 3D Reconstruction of Reflective, Transparent, and Low-Texture Objects},
author={Liang, Zhicheng and Yu, Haoyi and Li, Boyan and Zhang, Dayou and Cao, Zijian and Gong, Tianyi and Liu, Junhua and Cui, Shuguang and Wang, Fangxin},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={7244--7255},
year={2026}
}
License and Third-Party Material Notice
3DReflecNet is released under CC BY-NC 4.0 for non-commercial research and educational use.
Unless otherwise noted, this license applies to the dataset authors’ contributions, including rendered images, masks, depth maps, normal maps, camera parameters, metadata, annotations, documentation, and dataset organization.
The dataset may include or be derived from source 3D assets, materials, textures, or generated assets obtained from third-party sources or tools. Such third-party components may be subject to their own copyright, license terms, or usage restrictions. The CC BY-NC 4.0 license does not grant users rights beyond those held by the dataset authors.
Raw third-party assets are not redistributed unless redistribution rights are confirmed. Users are responsible for ensuring that their use of the dataset complies with applicable laws, regulations, and any relevant third-party terms. Commercial use is not permitted without prior written permission from the dataset authors and, where applicable, the relevant third-party rights holders.
- Downloads last month
- 1,966