SC-NeRF: NeRF-based Point Cloud Reconstruction using a Stationary Camera for Agricultural Applications

🧾 Overview

This dataset complements our paper on a stationary-camera-based NeRF framework for high-throughput 3D plant phenotyping. Traditional NeRF pipelines require a moving camera around a static object, which is impractical in automated indoor phenotyping environments. Our method enables 3D point cloud (PCD) reconstruction using a stationary camera and a rotating object, significantly simplifying the imaging setup.

The SC-NeRF dataset includes videos, extracted frames, COLMAP pose estimates, NeRF training outputs, and final 3D point clouds. It supports the community in replicating, extending, or evaluating methods for indoor phenotyping under controlled imaging conditions with minimal hardware requirements.

📁 Dataset Directory Structure

SC-NeRF/
├── raw/ # Raw videos and extracted frames
├── pre/ # Preprocessed COLMAP outputs and sparse point clouds
├── train/ # Trained NeRF models and checkpoints
└── pcd/ # Final reconstructed point clouds (10M points)

`raw/`

Contains high-resolution video clips (.MOV) and extracted keyframes for six objects:

Apricot
Banana
Maize Ear
Bell Pepper
Potted Plant-1
Potted Plant-2

Each object includes two recording modes:

_sc: stationary camera and rotating object (target configuration)
_gt: moving camera and stationary object (used as ground truth)

`pre/`

Includes COLMAP pose estimations and sparse point clouds generated using the ns-process-data pipeline in Nerfstudio. This folder transforms stationary camera captures into NeRF-compatible input by simulating camera motion.

`train/`

Stores trained NeRF models for each object using the nerfacto trainer in Nerfstudio.

`pcd/`

Final 10M-point cloud representations of each object were filtered, scaled using metric references, and aligned with ground-truth models for evaluation.

🌿 Object Categories

The dataset covers six agriculturally relevant objects with varying geometric complexity:

Apricot
Banana
Bell Pepper
Maize Ear
Crassula ovata (Potted Plant 1)
Haworthia sp. (Potted Plant 2)

Each object is reconstructed under two experimental protocols:

SC – Stationary Camera, Rotating Object (Primary Contribution)
GT – Ground Truth: Moving Camera, Stationary Object (Reference)

🎥 Imaging Protocol

Captured using an iPhone 13 Mini at 4K, 30fps
Videos trimmed to 30s, keyframes extracted at 4–5 FPS
ArUco markers and a ping pong ball (Ø = 0.04 m) used for pose and scale calibration
Processed using COLMAP and Nerfstudio

🧪 Applications

High-throughput plant phenotyping
AI training for 3D reconstruction
Point cloud alignment and evaluation
Hyperspectral and multimodal NeRF fusion

📜 License

CC-BY-NC-4.0
Creative Commons Attribution-NonCommercial 4.0 International License

🔖 Citation

If you use this dataset in your work, please cite:

@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} }