FudanCVL/APRS_dataset

APRS Dataset

Dataset Description

APRS (Active Panoramic Referring Segmentation) is a large-scale benchmark dataset for active perception in 360° panoramic environments. Unlike passive referring segmentation that processes static images, APRS requires agents to actively explore continuous panoramic scenes by adjusting viewing directions to seek and segment targets based on natural language instructions.

Dataset Summary

• 🎯 7,420 samples across 4,971 diverse panoramic scenes
• 🏠 Indoor and outdoor 360° environments
• 📐 Four types of spatial referring expressions:
  • Egocentric: First-person directional references (e.g., "look left to find...")
  • Unique-Attribute: Distinctive object features (e.g., "the red sofa")
  • Allocentric: Third-person spatial relations (e.g., "the chair near the window")
  • Multi-hop: Complex relational reasoning (e.g., "look to the left to find the bed, then find the lamp on the table next to it")

Data Fields

Each sample contains:

• Filename: Image filename (panoramic view)
• Img_W, Img_H, Aspect_Ratio: Image dimensions and aspect ratio
• Pt_X, Pt_Y: Target point pixel coordinates
• Pt_X_Norm, Pt_Y_Norm: Normalized point coordinates [0, 1]
• Box_X, Box_Y, Box_W, Box_H: Bounding box (x, y, width, height) in pixels
• Box_X_Norm, Box_Y_Norm, Box_W_Norm, Box_H_Norm: Normalized bounding box [0, 1]
• Pt_Theta, Pt_Phi: Target point spherical coordinates (degrees)
  • Theta (θ): Horizontal angle [-180°, 180°]
  • Phi (φ): Vertical angle [-90°, 90°]
• Box_Theta, Box_Phi: Bounding box center spherical coordinates (degrees)
• Description: Natural language referring expression
• Category: Spatial reference type
  • EGO: Egocentric (first-person directional)
  • ALLO: Allocentric (third-person spatial relations)
  • UNIQ: Unique attributes
  • MULT: Multi-hop reasoning

Usage

Load with APRS Dataset Class

from aprs import APRSDataset

# Load directly from HuggingFace Hub
dataset = APRSDataset.from_hub(repo_id="FudanCVL/APRS_dataset", split="train")

# Access samples
sample = dataset[0]
print(f"Instruction: {sample.instruction}")
print(f"Category: {sample.category}")
print(f"Initial view: θ={sample.init_theta:.1f}°, φ={sample.init_phi:.1f}°")
print(f"Target view: θ={sample.target_theta:.1f}°, φ={sample.target_phi:.1f}°")

# Load panoramic image (BGR numpy array)
image = sample.load_image()

# Get bounding box
box_pixels = sample.box_pixels()  # (x, y, w, h) in pixels

Interactive 360° Visualization

# Clone the official repository
git clone https://github.com/FudanCVL/APRS.git
cd APRS

# Install dependencies
pip install -e ".[viewer]"

# Launch interactive viewer directly from HuggingFace
python tools/viewer_360.py --hf --split test --index 0

# Or download dataset and view locally
python tools/viewer_360.py --root APRS_dataset --split test --index 0

Viewer Controls:

• 🖱️ Drag mouse to rotate view
• ⌨️ WASD or Arrow keys for navigation
• R to reset to initial view
• 🟩 Green box shows target region

Citation

If you use this dataset, please cite:

@article{tang2026seek,
  title={Seek to Segment: Active Perception for Panoramic Referring Segmentation},
  author={Tang, Song and Hu, Shuming and Shuai, Xincheng and Ding, Henghui and Jiang, Yu-Gang},
  journal={arXiv preprint arXiv:2607.02497},
  year={2026}
}

Contact

• Song Tang: tangsong322@gmail.com
• Henghui Ding: henghuiding@gmail.com

License

This dataset is released under the MIT License.

Links

• 📄 Paper: arXiv:2607.02497
• 🌐 Project Page: https://henghuiding.com/APRS/
• 💻 Code: https://github.com/FudanCVL/APRS