SDrAwberry

A scale-referenced long-sequence 3D reconstruction and plant-level canopy-volume estimation pipeline for strawberry field phenotyping.

Overview

SDrAwberry is a Python pipeline for synchronized multi-camera RGB sequences collected along strawberry beds. It prepares overlapping temporal batches from long image sequences, runs Depth-Anything-3 / DA3-style feed-forward multi-view reconstruction per batch, exports COLMAP-compatible sparse models and point clouds, and fuses per-batch submaps into a global point cloud.

Fusion is streamed and uses cross-batch COLMAP track correspondences, roll-constrained pose estimation for near-linear bed-aligned robot traversal, duplicate suppression, and voxel-hash fusion. In the current coordinate convention, roll-constrained alignment is implemented as a single-axis rotation around the reconstruction y-axis plus translation.

Downstream plant extraction estimates a support plane, computes signed plane-relative distances, applies HSV-guided plant candidate extraction, runs DBSCAN clustering, and rejects ground-like planar clusters. Plant-level canopy volumes are estimated by reconstructing watertight alpha-shape concave hull meshes. Metric interpretation requires reference-object-based scale recovery, such as known reference-object dimensions or an externally measured scale factor.

Installation

Conda + Editable Pip

conda create -n sdrawberry python=3.10 -y
conda activate sdrawberry
pip install -e .

Conda Environment File

conda env create -f environment.yml
conda activate sdrawberry

Pip

python -m venv .venv
source .venv/bin/activate
pip install -e .

Major dependencies:

• Python >= 3.10
• numpy
• scipy
• open3d
• opencv-python
• matplotlib
• pyyaml
• tqdm
• pycolmap
• trimesh
• torch, if required by your DA3 installation
• Depth-Anything-3 / DA3 CLI as an external dependency

DA3 must be installed separately according to its own license and installation instructions. This repository calls the external da3 command and does not vendor or redistribute DA3 source code.

Expected Data Layout

Input scene:

scene_dir/
  images/
    <camera_id_1>/
      rgb/
        rgb_<frame_id>.png
    <camera_id_2>/
      rgb/
        rgb_<frame_id>.png
    <camera_id_3>/
      rgb/
        rgb_<frame_id>.png

After batching and DA3 export:

scene_dir/
  batches/
    01/
      images/
        <frame_id>_<camera_id>.png
  sparse/
    01/
      cameras.bin
      images.bin
      points3D.bin
      points.ply

After fusion, clustering, and meshing:

scene_dir/
  merged/
    merged_global.ply
  clusters/
    cluster_0001.ply
    cluster_0002.ply
  meshes/
    cluster_0001/
      cluster_0001_m*.ply
      summary_concave.json

Quickstart

Batching and DA3:

python scripts/run_batching.py \
  --scene_dir /path/to/scene \
  --batch_size 10 \
  --overlap_frames 5 \
  --symlink \
  --da3_export_ply

Fusion:

python scripts/run_fusion.py \
  --base_dir /path/to/scene \
  --sparse_subfolder sparse \
  --ply points.ply \
  --out /path/to/scene/merged/merged_global.ply \
  --window_kdt 4 \
  --window_match 2 \
  --pre_voxel 0.003 \
  --voxel 0.002

Clustering:

python scripts/run_clustering.py \
  --input_pcd /path/to/scene/merged/merged_global.ply \
  --out_dir /path/to/scene/clusters \
  --dbscan_eps 0.015 \
  --dbscan_min_pts 120 \
  --plane_dist_threshold 0.005 \
  --z_axis_inverted

Concave hull:

python scripts/run_concave_hull.py \
  --in_pcd /path/to/scene/clusters/cluster_0001.ply \
  --out_dir /path/to/scene/meshes/cluster_0001 \
  --alpha_voxel 0.006 \
  --alpha_m_list "2,3,4,5,6,7,8,9,10,15,20" \
  --min_triangles 2000 \
  --fill_holes \
  --save_only_selected

Full pipeline:

python scripts/run_full_pipeline.py \
  --config configs/example_scene.yaml

Configuration

The configs/*.yaml files provide stage defaults and can be overridden by CLI flags:

• configs/batch.yaml
• configs/fusion.yaml
• configs/clustering.yaml
• configs/concave_hull.yaml
• configs/example_scene.yaml

For example:

python scripts/run_fusion.py --config configs/fusion.yaml --base_dir /path/to/scene --voxel 0.002

Dataset

The public dataset will be available at:

https://huggingface.co/datasets/Sycamorers/SDrAwberry

Status: Dataset upload in progress. The README will be updated once the dataset is fully available.

After download, place each scene under a local directory using the layout described above, for example:

/path/to/SDrAwberry_data/
  scene_001/
    images/
      <camera_id>/
        rgb/
          rgb_<frame_id>.png

Do not commit raw images, sparse models, point clouds, meshes, zip files, or dataset downloads to git.

Output Interpretation

DA3-style feed-forward reconstructions may be internally consistent but non-metric before reference-object scale recovery. Before interpreting volumes in physical units, provide known physical reference-object dimensions or an externally measured scale factor. Without scale recovery, volumes are reported in reconstruction units.

Limitations

• Outdoor illumination variation can affect reconstruction quality and HSV thresholds.
• Unsupervised clustering may over-segment a plant or merge connected canopies.
• DA3 memory usage depends on the number of views and image resolution.
• Absolute metric scale requires reference objects or external calibration.

Citation

If you use SDrAwberry in your research, please cite:

Huang, Z., Lee, W. S., Qin, R., Medeiros, H., Jeon, H., & Zhu, H. (2026). SDrAwberry: Scale-referenced Depth-Anything-3 long-sequence reconstruction for Strawberry canopy volume estimation. Computers and Electronics in Agriculture, 252, 112088. https://doi.org/10.1016/j.compag.2026.112088

BibTeX:

@article{huang2026sdrawberry,
  title = {SDrAwberry: Scale-referenced Depth-Anything-3 long-sequence reconstruction for Strawberry canopy volume estimation},
  author = {Huang, Zijing and Lee, Won Suk and Qin, Ruoyao and Medeiros, Henry and Jeon, Hongyoung and Zhu, Heping},
  journal = {Computers and Electronics in Agriculture},
  volume = {252},
  pages = {112088},
  year = {2026},
  doi = {10.1016/j.compag.2026.112088},
  url = {https://www.sciencedirect.com/science/article/abs/pii/S0168169926006836}
}

License

This repository is released under the MIT License. Third-party dependencies, DA3, and datasets may have separate licenses.

Contact

GitHub: https://github.com/Sycamorers