Skip to content
ST
ResearchCurrent

AIIRA 3D Reconstruction Pipeline for Maize Phenotyping

Converting drone- and phone-recorded RGB videos of maize tassels into structured 3D point-cloud outputs for agricultural phenotyping.

01 / Overview

I am working on an agricultural AI research project focused on converting drone- and phone-recorded RGB videos of maize tassels into 3D point-cloud outputs. The project supports maize phenotyping by using computer vision and 3D reconstruction methods to move from raw field video to structured spatial representations of plants.

02 / Technical focus

This project combines applied computer vision, 3D reconstruction, agricultural AI, field data collection, and research data engineering. A key part of the work is making the pipeline organized and reproducible across different recording sources, including drone and phone footage.

03 / Pipeline

  1. Raw RGB video01
  2. Frame extraction02
  3. COLMAP camera pose estimation03
  4. NeRF / Nerfstudio training04
  5. Point-cloud export05
  6. Tassel-level .ply extraction06

04 / What I worked on

  • Building computer vision pipelines for agricultural phenotyping using RGB video collected from field environments.
  • Processing drone and phone recordings into extracted frames for reconstruction workflows.
  • Supporting Structure-from-Motion pipelines using COLMAP for feature extraction, feature matching, geometric verification, triangulation, and camera pose estimation.
  • Running NeRF and Nerfstudio workflows to reconstruct 3D representations from 2D frames and estimated camera poses.
  • Organizing data-processing stages across raw video, preprocessing, pose estimation, NeRF training, point-cloud export, and individual tassel extraction.
  • Preparing point-cloud outputs and .ply files for downstream analysis of individual maize tassels.

05 / Tools and topics

PythonCOLMAPNerfstudioNeRFOpenCVOpen3DCloudCompareCUDAJetstream2LinuxStructure from Motion3D reconstructionPoint clouds.ply files