Plane detection and ranking via model information optimisation

Daoxin Zhong, Jun Li, Michael Chuah

[arXiv]

TLDR: Given an depth image with known intrinsics, find an ordered list of planes that represents the most likely model by minimising model information.

---

Data Format

root_dir/
├── rgb/
│   ├── 0.png
│   ├── 1.png
│   └── ...
├── depth/
│   ├── 0.png
│   ├── 1.png
│   └── ...

Get the NYU dataset and all predicted planes from here.

---

Docker

Build the docker image using the following command:

docker build \
    --ssh default=$SSH_AUTH_SOCK \
    -t info_opt .

Run the docker image using the following command:

docker run \
    -it \
    -v <path_of_scratchdata>:/scratchdata \
    --gpus all \
    --shm-size 16g \ 
    -d \
    --network=host \
    --restart unless-stopped \
    --env="DISPLAY" \
    --env="QT_X11_NO_MITSHM=1" \
    --volume="/tmp/.X11-unix:/tmp/.X11-unix:rw" \
    --device=/dev/ttyUSB0 \
    -e DISPLAY=unix$DISPLAY \ 
    --privileged \
    info_opt

---

Build

Python

git clone https://github.com/tauzn-clock/InformationOptimisation
pip3 install -r requirements.txt

C++

git clone https://github.com/tauzn-clock/InformationOptimisation
cd InformationOptimisation/cpp
chmod +x ./requirements.sh
./requirements.sh
mkdir build && cd build
cmake ..
make

The C++ implementation does not allow for image segmentation via SAM, so planes are found from the full depth image directly.

---

Run

Python

From python directory, run:

python3 demo.py ./nyu.yaml

The noise function can be changed at line 53 of python/demo.py.

C++

From cpp/build directory, run:

./main ../src/nyu.yaml

The noise function can be changed at line 93 of cpp/src/information_optimisation.cpp.

YAML Parameters:

• file_path: Path to data.

• img_count: Number of images to process.

• camera_params: Camera parameters, includes focal lengths (fx, fy), principal point (cx, cy).

• depth_max : Maximum depth value (in meters).

• resolution : Depth image resolution (in meters)

• conf: Confidence level for plane fitting.

• inlier_th: Assumed Inlier threshold for plane fitting.

• max_plane: Maximum number of models to test.

• use_sam: Use Segment Anything Model (SAM) for plane segmentation.

• sam_conf: Confidence level for plane fitting in each SAM region.

• sam_inlier_th: Assumed Inlier threshold for plane fitting in each SAM region.

• sam_max_plane: Maximum number of models to test in each SAM region.

---

Citation

@inproceedings{zhong2025infoopt,
  author = {Zhong,  Daoxin and Li,  Jun and Chuah,  Meng Yee Michael},
  title = {Plane Detection and Ranking via Model Information Optimization},
  booktitle = {Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  publisher = {IEEE},
  year = {2025},
  doi = {10.48550/ARXIV.2508.09625},
  url = {https://arxiv.org/abs/2508.09625},
}