DAL-PCQA

DAL-PCQA (Distortion-Aware Language-Annotated Point Cloud Quality Assessment Dataset) provides structured distortion annotations and natural language quality descriptions for point cloud quality assessment (PCQA).

Unlike traditional PCQA datasets that primarily provide Mean Opinion Scores (MOS), DAL-PCQA includes:

• Multi-level distortion severity labels
• 3D-specific geometric and structural distortion annotations
• 5-level quality labels: Excellent, Good, Fair, Poor, and Bad
• Structured natural language descriptions of perceived point cloud quality

The dataset is designed to support language-driven, multimodal, and explainable PCQA research.

---

Dataset Files

This Hugging Face repository contains annotation files derived from two source PCQA datasets:

• Annotations_SJTU.csv
• Annotations_WPC.csv
• Annotations_SJTU.parquet
• Annotations_WPC.parquet

The CSV files are provided for transparency and compatibility. The Parquet files are recommended for loading with Hugging Face Datasets because they preserve annotation labels such as None as strings rather than interpreting them as null values.

---

Dataset Splits

The dataset can be loaded as two splits:

• sjtu: annotations corresponding to the SJTU-PCQA dataset
• wpc: annotations corresponding to the WPC dataset

Example:

from datasets import load_dataset
repo_id = "swarnachakraborty96/DAL-PCQA"
data_files = {
    "sjtu": "Annotations_SJTU.parquet",
    "wpc": "Annotations_WPC.parquet"
}
dataset = load_dataset(repo_id, data_files=data_files)
print(dataset)
print(dataset["sjtu"][0])
print(dataset["wpc"][0])

---

Annotation Columns

Each row includes:

• Ply_name: point cloud file identifier
• Texture Condition: perceived texture quality
• Brightness distortion: brightness-related distortion severity
• Color Distortion: color-related distortion severity
• Noise: noise severity
• Blurriness: blur severity
• Point Density Distortion: point density distortion severity
• Scattering Artifact: scattering artifact severity
• Grid Artifact: grid artifact severity
• Missing Region: missing-region artifact severity
• Deformed Shape: shape deformation severity
• MOS: original Mean Opinion Score
• Rounded MOS: rounded MOS value
• Label: 5-level quality label
• Generated_Description: structured natural language quality description

---

Original Point Cloud Files

This repository does not include the original point cloud files.

To obtain the point clouds, please download them from the official source dataset repositories:

• SJTU-PCQA: Paper | Dataset
• WPC Dataset: Paper | Dataset

The Ply_name column in DAL-PCQA corresponds to point cloud file names from these source datasets.

---

Code and Evaluation Scripts

This Hugging Face repository contains the DAL-PCQA annotation files only.

The evaluation scripts and model-specific testing code are available in the project GitHub repository:

GitHub Repository

The GitHub repository includes scripts for evaluating the dataset with different multimodal and vision-language models.

---

Associated Paper

This dataset is associated with the following paper:

DAL-PCQA: Distortion-Aware Language-Annotated Point Cloud Quality Assessment Dataset

Paper page / arXiv link: https://arxiv.org/abs/2606.07938

If you use this dataset, please cite the associated paper.

@article{chakraborty2026dal,
  title={DAL-PCQA: Enabling Distortion-Level and Language-Driven Reasoning for Point Cloud Quality Assessment},
  author={Chakraborty, Swarna and Ara{\~A}{\=e}jo, Gabriel De Castro and Faria, Syeda Tasmi and Carvalho, Marcelo M and Farias, Mylene CQ},
  journal={arXiv preprint arXiv:2606.07938},
  year={2026}
}

---