IITKGP_Fence dataset

Overview

The IITKGP_Fence dataset is designed for tasks related to fence-like occlusion detection, defocus blur, depth mapping, and object segmentation. The captured data vaies in scene composition, background defocus, and object occlusions. The dataset comprises both labeled and unlabeled data, as well as additional video and RGB-D data. The contains ground truth occlusion masks (GT) for the corresponding images. We created the ground truth occlusion labels in a semi-automatic way with user interaction.

Dataset Structure

Click to expand directory structure

└───IITKGP_Fence dataset
    ├───Labeled
    │   ├───GT
    │   │       (1).png
    │   │       ...
    │   │       (175).png
    │   │       
    │   ├───Imgs
    │   │       (1).png
    │   │       ...
    │   │       (175).png
    │   │       
    │   └───Others
    │       ├───Bird
    │       │       c1.png
    │       │       ...
    │       │       f1.png
    │       │       
    │       ├───Blurred_bg
    │       │   ├───GT
    │       │   │       gt1.jpg
    │       │   │       ...
    │       │   │       gt88.jpg
    │       │   │              
    │       │   └───Imgs
    │       │       im (1).jpg
    │       │       ...
    │       │       im (88).jpg
    │       │                      
    │       └───...
    └───Unlabeled
        ├───DefocusBlurred
        │   └───Blurred_fg
        │       │   BlMov_01.mov
        │       │   ...
        │       │   BlMov_46.mov
        │       │   
        │       └───Imgs
        │               BlJPG_01.jpg
        │               ...
        ├───RGB
        │   ├───Imgs
        │   │       BlJPG_001.jpg
        │   │       ...
        │   │       BlJPG_205.jpg
        │   │       
        │   └───Vids
        │       ├───utils
        │       └───Zoo
        │               Zoo_001.mp4
        │               ...
        │               Zoo_205.mp4
        └───RGBD
            │   data1.mat
            │   ...   
            ├───Othersamples
            │   ├───DATA1
            │   └───...
            └───utils

Dataset Description

Here's an overview of its structure and contents:

Labeled Data:
- GT (Ground Truth): Contains 175 PNG images representing the ground truth labels for corresponding input images.
- Imgs: Contains 175 PNG images, which are RGB images that correspond to the ground truth.
- Others:
  - Multiple Scenes: Contains various scenes (e.g., Bird/, Tennis/, etc.). Each scene consists of four pairs of RGB images and their corresponding ground truth masks.
  - Blurred_bg: This folder includes images with blurred backgrounds and corresponding ground truth occlusion segmentation labels.
    - GT: Contains 88 ground truth occlusion mask.
    - Imgs: Contains 88 blurred background images in JPG format.
Unlabeled Data:
- DefocusBlurred: Focused on data related to blurred foreground occlusions.
  - Blurred_fg: Contains 46 video files and `Imgs/'.
- RGB: Contains regular RGB images and videos.
  - Imgs: 205 JPEG images of various scenes with occlusions.
  - Vids: Includes a total of 214 video files.
- RGBD: Contains data for scenes with RGB images and depth data.
  - MAT files: These files store all the data values and additional camera information for various samples.
  - Othersamples: Includes additional data samples captured in laboratory.

Key Dataset Features:

Fence Detection: Designed for detecting fences or fence-like structures that might occlude objects.
Defocus Blur: Contains images and videos with blurred objects, likely to challenge detection and segmentation algorithms.
RGBD Data: Offers depth information alongside RGB images, which can be used for tasks like 3D reconstruction or occlusion handling.
Unlabeled and Labeled Data: Facilitates both supervised and unsupervised learning tasks. The Labeled folder data provides ground truth occlusion masks, while the Unlabeled folder data allows for further experimentation or self-supervised methods.

Dataset Repository

GitHub Repository: Occlusion-Removal
Paper: Deep Generative Adversarial Network for Occlusion Removal from a Single Image
- Authors: Sankaraganesh Jonna, Moushumi Medhi, Rajiv Ranjan Sahay

Usage

pip install datasets

from datasets import load_dataset
dataset = load_dataset('NeuroVizv0yaZ3R/IITKGP_Fence_dataset')

Citation

BibTeX:

@misc{jonna2024deepgenerativeadversarialnetwork,
      title={Deep Generative Adversarial Network for Occlusion Removal from a Single Image}, 
      author={Sankaraganesh Jonna and Moushumi Medhi and Rajiv Ranjan Sahay},
      year={2024},
      eprint={2409.13242},
      archivePrefix={arXiv},
      url={https://arxiv.org/abs/2409.13242}, 
}

License

For license details, refer to the LICENSE file.

Acknowledgments

We would like to express our gratitude to Dr. Sreeja S R for her valuable contributions to the Imgs folder within the RGB directory of the Unlabeled dataset, which greatly enriched its diversity and quality.

Contact

medhi.moushumi@iitkgp.ac.in

image imagewidth (px) 320 5.31k