README.md

---
pretty_name: Transparent Object Point Cloud Dataset


language:
  - en
license: other
task_categories:
  - image-to-3d
tags:
  - transparent-objects
  - point-cloud
  - 6d-pose-estimation
  - registration
  - icp
  - rgb-d
size_categories:
  - 1K<n<10K
---

# Transparent Object Point Cloud Dataset

[Project Page](https://viugic.github.io/) |
[Paper]() |
[Code]() |
[Dataset](https://huggingface.co/datasets/adfa5456/Final_Dataset)


## Dataset Description

This dataset contains multi-view RGB images, raw and processed point clouds, object-level reference point clouds, and 6D pose annotations for transparent-object pose estimation and point-cloud registration.

The dataset is organized into two subsets:

- **Printed objects:** 6 object instances (`Eyeglass`, `Kettle`, `Lightbulb`, `Lighter`, `Magnifying_glass`, and `Spray`).
- **Scanned objects:** 9 object instances (`bulb_0`, `bulb_1`, `cctv_0`, `cctv_1`, `cup_0`, `jar_0`, `jar_1`, `pet_0`, and `pet_1`).

The current repository contains approximately 1,500 views in total. Most objects have about 100 views. The data occupies approximately 73 GB when stored uncompressed on disk.



## Intended Uses

The dataset may be useful for research on:

- Transparent-object 6D pose estimation
- Point-cloud registration and ICP
- Depth or point-cloud completion
- RGB/point-cloud multimodal reconstruction
- Robust pose estimation under missing or noisy depth measurements

This dataset has not been validated for safety-critical, medical, or production robotic applications.

## Dataset Structure

```text
.
├── dataset_printed/
│   ├── Image/<object>/<view>.png
│   ├── raw/<object>/<view>.ply
│   ├── processed/<object>/noisy_filtered_<view>.ply
│   ├── gt/<object>/noisy_filtered_<view>.json
│   └── gt_pcd/<object>/gt_filtered.ply
└── dataset_scanned/
    ├── Image/<object>/<view>.png
    ├── raw/<object>/<view>.ply
    ├── processed/<object>/noisy_filtered_<view>.ply
    ├── gt/<object>/result_<view>.txt
    ├── gt_pcd/<object>/gt_filtered.ply
    └── vggt/<object>/
        ├── <view>.ply
        └── <view>_conf.txt
```

There are no predefined train, validation, or test splits. Users should define object-wise or view-wise splits appropriate for their experiments and report them with their results.

### File Types

| Path | Format | Description |
|---|---|---|
| `Image/` | PNG | RGB image associated with each view. Images are 1932 x 1096 pixels in the current release. |
| `raw/` | PLY | Raw point cloud captured for a view. |
| `processed/` | PLY | Filtered/processed version of the corresponding raw point cloud. |
| `gt/` | JSON or TXT | Ground-truth 6D pose. Printed objects use JSON; scanned objects use a 4 x 4 matrix stored as text. |
| `gt_pcd/` | PLY | Reference point cloud for each object instance. |


### View Naming Convention

Each sample follows the naming convention:

```text
<fill_rate>_<sample_index>
```

## Loading the Dataset

This repository stores the original files directly and does not currently provide a Hugging Face `datasets` loading script. Download a snapshot while preserving the directory structure:

```python
from huggingface_hub import snapshot_download

dataset_dir = snapshot_download(
    repo_id="adfa5456/Final_Dataset",
    repo_type="dataset",
)
print(dataset_dir)
```

Example: load an RGB image, a point cloud, and a scanned-object pose:

```python
from pathlib import Path

import numpy as np
import open3d as o3d
from PIL import Image

root = Path(dataset_dir)
object_name = "pet_1"
view = "0_1"

image = Image.open(root / "dataset_scanned" / "Image" / object_name / f"{view}.png")
raw_pcd = o3d.io.read_point_cloud(
    str(root / "dataset_scanned" / "raw" / object_name / f"{view}.ply")
)
processed_pcd = o3d.io.read_point_cloud(
    str(root / "dataset_scanned" / "processed" / object_name / f"noisy_filtered_{view}.ply")
)
pose = np.loadtxt(root / "dataset_scanned" / "gt" / object_name / f"result_{view}.txt")

print(image.size, np.asarray(raw_pcd.points).shape, pose.shape)
```

Install the example dependencies with:

```bash
pip install huggingface_hub numpy open3d pillow
```

For large repositories, downloading only selected files with `huggingface_hub.hf_hub_download` or CLI include/exclude patterns may be more practical than downloading the entire dataset.

## Personal and Sensitive Information

The dataset is intended to contain tabletop object observations only.

Permission is required for redistribution, modification, or commercial use.

For permission requests, please contact [keep9642@korea.ac.kr].

## License

Copyright (c) 2026 RILAB and SNUAILAB.

This dataset was created and maintained jointly by:

- [RILAB](https://sites.google.com/view/sungjoon-choi/home)
- [SNUAILAB](https://snuailab.ai/)

The YAML metadata currently uses `license: other` as a placeholder. Do not publish the dataset card without replacing or explaining this value.

## Citation

If you use this dataset, please cite:

```bibtex
@inproceedings{park2026viugic,
  title  = {Pose Estimation of Transparent Objects via Depth Completion and Confidence-Guided Registration},
  author = {Jeongeun Park, Yeoncheol Jang, Changjin Kim, YoungJoon Yoo, Sungjoon Choi},
  year   = {2026}
}

```