datasets_cards.yaml

Dataset Metadata

dataset_info:
  name: SC-NeRF
  description: >
    SC-NeRF is a dataset designed for 3D reconstruction using Neural Radiance Fields (NeRF) under a 
    stationary-camera setup. It targets high-throughput plant phenotyping in controlled indoor environments, 
    simplifying the traditional NeRF pipeline that requires a moving camera around static objects. 
    Instead, it uses a rotating object in front of a stationary camera, making it practical for automated 
    phenotyping systems. The dataset includes videos, extracted frames, COLMAP pose estimations, trained NeRF 
    models, and high-resolution point clouds for six agriculturally relevant objects.

  version: 1.0
  license: CC-BY-NC-4.0
  authors:
    - Kibon Ku
    - Talukder Z. Jubery
    - Elijah Rodriguez
    - Aditya Balu
    - Soumik Sarkar
    - Adarsh Krishnamurthy
    - Baskar Ganapathysubramanian

  citation: >
    @article{ku2025stationarynerf,
      title = {NeRF-based Point Cloud Reconstruction using a Stationary Camera for Agricultural Applications},
      author = {Kibon Ku, Talukder Z. Jubery, Elijah Rodriguez, Aditya Balu, Soumik Sarkar, Adarsh Krishnamurthy, Baskar Ganapathysubramanian},
      year = {2025},
      journal = {arXiv preprint arXiv:2503.21958}
    }

  intended_use:
    - Stationary-camera-based 3D reconstruction
    - High-throughput plant phenotyping
    - AI-based point cloud generation
    - Benchmarking indoor NeRF pipelines
    - Hyperspectral and multimodal NeRF fusion

  features:
    - Videos (.MOV)
    - Keyframes (JPG/PNG)
    - COLMAP outputs (poses, sparse PCD)
    - Trained NeRF models (nerfacto, Nerfstudio format)
    - Final reconstructed 10M-point point clouds (.ply)

  dataset_size:
    raw:
      - "6 video objects × 2 capture types (SC and GT) in .MOV format"
      - "Keyframes extracted at 4–5 FPS per object"
    pre:
      - "COLMAP pose estimates and sparse point clouds for all objects"
    train:
      - "Nerfstudio-trained NeRF models with checkpoints"
    pcd:
      - "Final 10M-point point clouds for 6 objects (SC and GT), aligned and filtered"

  dependencies:
    - Python 3.8+
    - Nerfstudio (https://docs.nerf.studio)
    - COLMAP
    - Open3D (for visualization and evaluation)
    - CloudCompare or MeshLab (optional for inspection)

  installation_instructions: |
    Clone and set up the dataset locally:
    ```bash
    git clone https://huggingface.co/datasets/BGLab/SC-NeRF
    cd AgriPCD
    ```

  download_instructions: |
    1. Download the dataset files from the Hugging Face repository or provided links.
    2. Unzip the folders:
       ```bash
       unzip raw.zip
       unzip pre.zip
       unzip train.zip
       unzip pcd.zip
       ```

  training_instructions: |
    Preprocess and train NeRF models using Nerfstudio:
    ```bash
    ns-process-data --data ./pre/object_name
    ns-train nerfacto --data ./pre/object_name
    ```

  pointcloud_extraction: |
    Export the high-resolution point cloud:
    ```bash
    ns-export pointcloud --load-config ./train/object_name/config.yml
    ```

  evaluation_instructions: |
    Align reconstructed and ground truth point clouds using ICP and evaluate using precision/recall or other geometric metrics.

  visualization_instructions: |
    Visualize point clouds using Open3D:
    ```python
    import open3d as o3d
    pcd = o3d.io.read_point_cloud("pcd/apricot_sc.ply")
    o3d.visualization.draw_geometries([pcd])
    ```

  repository_links:
    - https://huggingface.co/datasets/BGLab/SC-NeRF
    - https://arxiv.org/abs/2503.21958
    - https://docs.nerf.studio