Datasets:
Acknowledge the COGS Dataset License
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Please read and agree to the COGS Dataset License v1.0 before accessing the dataset. License: https://huggingface.co/datasets/XiaoZhengCV/COGS/blob/main/LICENSE
Log in or Sign Up to review the conditions and access this dataset content.
COGS: A Diagnostic Benchmark for Attribute-Aware Object-Level 3D Gaussian Splatting Perception
COGS (Common Objects of 3D Gaussian Splatting) is an object-centric 3D Gaussian Splatting (3DGS) dataset designed for systematic analysis of Gaussian attribute utilization in 3D perception.
COGS contains 15,675 curated 3DGS objects across 40 categories, with appearance variation, artifact-filtered structures, and position, opacity, scale, rotation, and DC spherical-harmonic appearance attributes.
COGS is introduced alongside COGS-Bench, a benchmark for object classification and attribute-usage diagnosis, and HALO, an attribute-aware reference model.
Paper status: Accepted at the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2027.
Dataset Overview
| Property | Description |
|---|---|
| Dataset | COGS |
| Representation | Object-level 3D Gaussian Splatting |
| Objects | 15,675 |
| Categories | 40 |
| Available versions | Full-resolution and FPS-1024 |
| File format | Binary little-endian PLY, with float32 vertex properties |
| Primary task | 40-class object classification |
| Additional purpose | Gaussian attribute-usage diagnosis |
| Full-resolution size | Approximately 449.98 GB of uncompressed PLY files |
| FPS-1024 size | Approximately 1.10 GB of uncompressed PLY files |
| License | COGS Dataset License v1.0 |
Dataset Construction
COGS is constructed from publicly sourced object images through a multi-stage curation and generative reconstruction pipeline:
- Image collection and pre-filtering.
- Semantic localization with GroundingDINO.
- Foreground-aware filtering, CLIP-based deduplication, and manual image inspection.
- Generative 3DGS reconstruction using TRELLIS.
- Interactive 3DGS quality inspection to remove severe structural artifacts.
Only the final 3DGS assets and associated benchmark data are distributed.
Dataset Versions
Full-resolution
The complete curated reconstructed 3DGS assets, before the processing used to create the fixed FPS-1024 version. Objects contain between 18,208 and 2,425,408 Gaussian primitives, with a median of 355,520.
Each asset contains P/O/S/Q/C attributes as specified below. Appearance is represented by DC spherical-harmonic coefficients; higher-order SH coefficients are not present.
FPS-1024
A fixed processed version containing 1,024 Gaussian primitives per object. Its preparation includes FPS, rotations about the Y-axis followed by the X-axis, and random partial removal of Gaussian primitives.
For reproducible classification benchmarking, use the provided FPS-1024 files rather than regenerating them from the full-resolution assets. The rotations and partial removal are already reflected in these files and should not be repeated as an additional offline preparation step. Training-time augmentation is a separate part of the experiment configuration.
File Organization
The following shows the extracted layout used by the current data and loader. Public archive names will be specified when packaging is finalized.
COGS/
├── README.md
├── LICENSE
├── gs/
│ ├── 00_apple/
│ │ ├── 0000.ply
│ │ └── ...
│ ├── ...
│ └── 39_shelf/
└── gs_fps_1024/
├── categories.txt
├── train_benchmark.json
├── test_benchmark.json
├── train_pretrain.json
├── test_pretrain.json
├── 00_apple/
│ ├── 0000.ply
│ └── ...
├── ...
└── 39_shelf/
categories.txt maps integer labels 0-39 to category names. Split JSON files contain lists of object paths relative to the corresponding data root, such as 00_apple/0000.ply. The same relative object paths identify the corresponding full-resolution assets under gs/.
Keep the category file and all four split files alongside the FPS category directories. They are needed by the provided classification loader.
Data Format
Each PLY contains a vertex element with 17 float32 properties:
| Attribute | PLY properties | Stored representation and interpretation |
|---|---|---|
| Position (P) | x, y, z |
Gaussian center coordinates |
| Opacity (O) | opacity |
Logit opacity; apply sigmoid to obtain opacity |
| Scale (S) | scale_0, scale_1, scale_2 |
Log scale; apply exp to obtain positive scales |
| Rotation (Q) | rot_0, rot_1, rot_2, rot_3 |
Quaternion in w, x, y, z order; normalize before use |
| Appearance (C) | f_dc_0, f_dc_1, f_dc_2 |
DC spherical-harmonic coefficients, not directly RGB values |
| Normal placeholders | nx, ny, nz |
All zero; these are not valid surface normals |
Limitations
- Generative geometry: COGS contains image-conditioned generated 3DGS assets. It should not be treated as ground-truth geometry from physical measurements or multi-view capture.
- Residual artifacts: Manual filtering reduces severe structural defects but does not guarantee that every asset is artifact-free or physically accurate.
- Coverage and imbalance: The dataset covers 40 object categories with unequal category sizes. It does not represent all object types, environments, or real-world scene distributions.
- Representation: Appearance is stored using DC SH coefficients only; the zero normal fields do not provide surface-normal supervision.
Access and License
COGS is distributed under the COGS Dataset License v1.0 for non-commercial research and educational purposes. Users must agree to its terms before accessing the dataset.
Research colleagues and collaborators may share the dataset privately under the conditions specified in the license. Public redistribution, resale, or hosting requires prior written authorization. See LICENSE for the complete terms, including attribution, permitted research illustrations, termination, and rights-holder removal requests.
The accompanying software code is distributed separately under the MIT License.
Access and Download Workflow
The intended release uses automatic approval after the access form is submitted:
- Sign in to Hugging Face and open the COGS dataset repository.
- Read the license, provide the requested name and affiliation, and acknowledge the license terms.
- Once access is granted, use the repository's Files tab to download the desired version and the category/split metadata. FPS-1024 is sufficient for the classification benchmark; full-resolution data is a separate, much larger download.
Code and Contact
Public code repository: COGS
For dataset questions or rights-holder removal requests, use the official repository's Community tab or contact xiaozheng19@mails.ucas.ac.cn. Removal requests should identify the affected object paths and provide supporting information, as described in the license.
Citation
If you use COGS in your research, please cite the following accepted paper:
@inproceedings{xiao2027cogs,
title = {COGS: A Diagnostic Benchmark for Attribute-Aware Object-Level 3D Gaussian Splatting Perception},
author = {Xiao, Zheng and Liu, Yifan and Deng, Weijian and Li, Jiashu and Liu, Ziqin and Han, Zhenjun and Jiao, Jianbin},
booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
year = {2027},
note = {Accepted for publication}
}
The citation will be updated with final proceedings details when available.
- Downloads last month
- 43