PhenoIMC3D: Iterative Motion Compensation for Canonical 3D Reconstruction (RA-L 2026)

Andre Rochow*, Jonas Marcic*, Svetlana Seliunina, and Sven Behnke
* equal contribution

[Paper] | [arXiv] | [Dataset]

---

Without Motion Compensation	PhenoIMC3D (ours)

---

Description

Official implementation of Iterative Motion Compensation for Canonical 3D Reconstruction From UAV Plant Images Captured in Windy Conditions (accepted to IEEE Robotics and Automation Letters, 2026) by Andre Rochow*, Jonas Marcic*, Svetlana Seliunina, and Sven Behnke.

The method models and compensates unwanted motion in scenes captured by UAVs, such as plant movement caused by downwash. These effects can significantly degrade the quality of 3D reconstruction.

PhenoIMC3D enables:

• motion compensation in dynamic scenes
• extraction of a single sharp canonical 3D representation
• compatibility with various 3D reconstruction methods

PhenoIMC3D can be seamlessly combined with:

• NeRF-based methods
• 3D Gaussian Splatting
• other recent splatting-based methods

To integrate PhenoIMC3D with your reconstruction pipeline, follow the instructions in the Setup and Configuration sections.

---

Setup

Choose and install a 3D reconstruction method of your choice.

We provide examples for:

• 3D Gaussian Splatting
• Deformable Beta Splatting

However, any other reconstruction method can also be used.

Clone the repository

git clone https://github.com/AIS-Bonn/PhenoIMC3D.git --recursive
cd PhenoIMC3D/submodules/RAFT
./download_models.sh

Install dependencies

Activate the environment of your selected reconstruction method:

conda activate your_3d_method

Install required packages (if not already available in your environment):

pip install torch torchvision
pip install scipy
pip install opencv-python
pip install tqdm
pip install natsort

---

Configuration

We provide configuration files for integrating PhenoIMC3D with:

• 3D Gaussian Splatting
• Deformable Beta Splatting

You can also define configurations for custom reconstruction methods.

Placeholders

Placeholder	Description
${dataset_path}	Path to the dataset (provided via CLI arguments)
${model_path}	Path to the output model/workspace directory

These placeholders are automatically replaced at runtime.

---

3D Gaussian Splatting

configs/gaussian_splatting.json

{
    "train_cmd": "python /path/to/your/installed/gaussian-splatting/train.py -s ${dataset_path} -m ${model_path} -r 1",
    "model_weights": "${model_path}/point_cloud/iteration_30000/point_cloud.ply",
    "render_cmd": "python /path/to/your/installed/gaussian-splatting/render.py -m ${model_path} --skip_test",
    "img_predictions": "${model_path}/train/ours_30000/renders"
}

---

Deformable Beta Splatting

configs/beta_splatting.json

{
    "train_cmd": "python /path/to/your/installed/beta-splatting/train.py -s ${dataset_path} --model_path ${model_path} --resolution 1 --iterations 30000 --disable_viewer",
    "model_weights": "${model_path}/point_cloud/iteration_30000/point_cloud.ply",
    "render_cmd": "python /path/to/your/installed/beta-splatting/render.py -m ${model_path} --skip_test",
    "img_predictions": "${model_path}/train/ours_30000/renders"
}

---

Custom 3D Reconstruction Method

{
    "train_cmd": "COMMAND TO RUN TRAINING OF YOUR RECONSTRUCTION METHOD",
    "model_weights": "OPTIONAL: PATH TO MODEL CHECKPOINTS",
    "render_cmd": "COMMAND TO RUN RENDERING",
    "img_predictions": "PATH TO PREDICTED IMAGES"
}

---

Prepare Dataset

To use our dataset, download it via the link above. It contains high-resolution images (8064×6048), so you may want to downscale them before training. A script is provided for this purpose. Note that, due to RAFT, both height and width must be divisible by 8.

python scale_dataset.py --dataset_path /path/to/datset/scene --colmap_binary /path/to/your/colmap/binary --r resolution

In our paper, we used --r 4, corresponding to a resolution of (2016×1512).

---

Training

3D Gaussian Splatting

conda activate gaussian_splatting

python train.py \
    --config configs/gaussian_splatting.json \
    --dataset_path /path/to/dataset/scene \
    --model_path output/workspace/model_path

---

Deformable Beta Splatting

conda activate beta_splatting

python train.py \
    --config configs/beta_splatting.json \
    --dataset_path /path/to/dataset/scene \
    --model_path output/workspace/model_path

---

Some Results

Without Motion Compensation	PhenoIMC3D (ours)

For additional results and a quantitative evaluation, please refer to our paper.

---

Citation

If you use this work, please cite:

@article{rochow2026iterative,
  title={Iterative Motion Compensation for Canonical 3D Reconstruction From UAV Plant Images Captured in Windy Conditions},
  author={Rochow, Andre and Marcic, Jonas and Seliunina, Svetlana and Behnke, Sven},
  journal={IEEE Robotics and Automation Letters},
  year={2026},
  publisher={IEEE}
}