Door Identification and Pose Estimation Dataset

This repository contains the dataset used for the paper "Identification and Pose Estimation for Doors Under Occlusion Using Synthetic and Real Data".

Overview

The dataset is designed for training and evaluating door detection and 6-DoF pose estimation models in indoor environments. The data includes both synthetic and real-world images of various door types in different configurations, lighting conditions, and levels of occlusion.

Our dataset can be used for a variety of tasks, including:

• Door detection
• Pose estimation (6-DoF)
• Training models for handling occluded doors
• Enhancing the generalization of models using synthetic data

Contents

• Real-Dataset: 4,000 images (3,000 from a public dataset and 1,000 captured manually by our team).
• Synthetic Dataset (NVSII): 50,000 automatically generated images using the NVSII method, with various occlusion and lighting conditions.
• Our Synthetic Dataset: 25,000 Full HD images automatically generated using the Isaac method and Blender for creating diverse 3D models and random occlusion scenarios.

Each dataset contains labeled data for both door detection and keypoint annotations to facilitate 6-DoF pose estimation.

Directory Structure

datasets/
├── real-dataset/
│   ├── images/              # Real-world RGB images
│   └── annotations/         # Corresponding keypoint annotations
├── synthetic-nvsii-dataset/
│   ├── images/              # Synthetic RGB images from NVSII
│   └── annotations/         # Keypoint annotations
├── synthetic-isaac-dataset/
│   ├── images/              # Synthetic RGB images from Isaac
│   └── annotations/         # Keypoint annotations
models/
├── yolov8/                  # YOLOv8 model configuration and weights
├── dope/                    # DOPE model configuration and weights
├── Gen6D/                   # Gen6D model configuration and tests
scripts/
├── train.py                 # Script for training the model
├── eval.py                  # Script for evaluating the model
└── dataset-prep.py          # Helper script for dataset preprocessing

Usage

1. Clone the Repository

git clone https://github.com/username/door-pose-estimation-dataset.git
cd door-pose-estimation-dataset

Download Instructions

The dataset is not included directly in this repository due to its size. You can request access to the dataset by contacting Renan Moreira at [email protected] or through the link provided in the paper after acceptance.

License

This dataset is made available for research purposes only. By using this dataset, you agree to cite the original paper and give proper credit to the authors.

Citation

If you use this dataset or the models in your research, please cite our paper:

@article{Moreira2024DoorDetection,
  title={Identification and Pose Estimation for Doors Under Occlusion Using Synthetic and Real Data},
  author={Renan Moreira, Thiago Segreto, Juliano D. Negri, João C. V. Soares, Marcelo Becker, and Vivian S. Medeiros},
  journal={To be published},
  year={2024}
}