Dataset Documentation

Overview

This dataset supports research in 3D reconstruction and computer vision, focusing on plant-based scenarios. It consists of two main components:

1. Captured images with corresponding poses
2. Ground truth LiDAR scans for evaluation

Research

This is used in the paper : Evaluating Neural Radiance Fields for 3D Plant Geometry Reconstruction in Field Conditions Available Online: https://spj.science.org/doi/full/10.34133/plantphenomics.0235

Dataset Structure

The dataset is organized into two primary folders:

1. Images and Poses
2. Ground Truth Point Clouds And Alignment Matrices

Images and Poses

This folder contains datasets captured using Polycam, processed for use with 'nerfstudio'. Each subfolder includes:

• Original images
• Downsampled images (by factors of 2, 4, and 8)
• A 'transform.json' file containing camera pose information

Ground Truth Point Clouds And Alignment Matrices

This folder contains the corresponding LiDAR scans (.ply files) that serve as ground truth for the captured scenes.

Scenario Identifiers

The dataset includes three main scenarios, each corresponding to a specific research setup:

1. CCL-scanned-data-multiple-polycam-images-processed:

   • Scenario-III in the research paper
   • Represents multiple plants outdoors in a field setting

2. CCL-scanned-data-outdoor-stage2-low-height-processed:

   • Scenario-II in the research paper
   • Represents multiple plants indoors

3. CCL-scanned-data-single-img-50-qual-90-processed:

   • Scenario-I in the research paper
   • Represents a single plant indoors

Additional Data

Folders prefixed with "lpips" (e.g., lpips-validation-data-processed) contain only images and poses, without corresponding ground truth data. These are used for additional validation or testing purposes.

Detailed Directory Structure

For those interested in the complete directory structure:

Dataset
├── Ground Truth Point Clouds And Alignment Matrices
│   ├── CCL-scanned-data-multiple-polycam-images-processed
│   ├── CCL-scanned-data-outdoor-stage2-low-height-processed
│   └── CCL-scanned-data-single-img-50-qual-90-processed
└── Images and Poses
    ├── CCL-scanned-data-multiple-polycam-images-processed
    ├── CCL-scanned-data-outdoor-stage2-low-height-processed
    ├── CCL-scanned-data-single-img-50-qual-90-processed
    ├── lpips-validation-data-processed
    ├── lpips-validation-data-processed-1
    ├── lpips-validation-data-processed-2
    ├── lpips-validation-data-processed-3
    ├── lpips-validation-data-processed-4
    └── lpips-validation-data-processed-5

Folder Contents Example

Here's an example of what you might find in a typical scenario folder:

CCL-scanned-data-single-img-50-qual-90-processed
├── evaluations
├── CCL-scanned-data-single-img-50-qual-90-processed_COLMAP_SfM.log
├── CCL-scanned-data-single-img-50-qual-90-processed_trans.txt
├── CCL-scanned-data-single-img-50-qual-90-processed-cleaned-old-v1.ply
├── CCL-scanned-data-single-img-50-qual-90-processed-uncleaned.ply
├── CCL-scanned-data-single-img-50-qual-90-processed.json
└── CCL-scanned-data-single-img-50-qual-90-processed.ply

Key files in this structure include:

• _trans.txt: Contains the alignment matrix
• .ply files: Various versions of the LiDAR scan, including cleaned and uncleaned versions
• .json file: (Optional) Used for evaluation scripts based on the 'Tanks And Temples repository'

Usage Notes

• The ground truth LiDAR scans are crucial for evaluating the accuracy of 3D reconstructions generated from the image datasets.
• When working with a specific scenario, ensure you use the corresponding images, poses, and ground truth data from the matching folders in both main directories.
• The lpips-prefixed folders can be used for additional validation but lack ground truth data.

This dataset structure allows for comprehensive research in 3D reconstruction techniques, particularly in plant-based scenarios, providing both the necessary input data (images and poses) and ground truth data for evaluation.

If you find this helpful, consider citing:

@article{arshad2024evaluating,
  title={Evaluating Neural Radiance Fields for 3D Plant Geometry Reconstruction in Field Conditions},
  author={Arshad, Muhammad Arbab and Jubery, Talukder and Afful, James and Jignasu, Anushrut and Balu, Aditya and Ganapathysubramanian, Baskar and Sarkar, Soumik and Krishnamurthy, Adarsh},
  journal={Plant Phenomics},
  volume={6},
  pages={0235},
  year={2024},
  publisher={AAAS}
}