Awesome

Polygon Detection for Room Layout Estimation using Heterogenous Graphs and Wireframes

This repo contains code for training and inference of our model for connected polygon detection in RGB images. See the paper on arXiv titled "Polygon Detection for Room Layout Estimation using Heterogenous Graphs and Wireframes" for more details.

Conceptual image of how the model works with wireframes and polyons.

Clone repo and get submodules

git clone --recurse-submodules git@github.com:DavidGillsjo/polygon-HGT.git

alternatively

git clone git@github.com:DavidGillsjo/polygon-HGT.git
git submodule init
git submodule update

Prerequisites

To run the code you need an NVIDIA GPU with their driver installed. Furthermore you need to install docker and gpu support for docker with nvidia container toolkit. [Link to docker install] [Link to nvidia container toolkit]

Docker

We supply a Dockerfile to build a docker image which can run the code. First, modify line 7 so that gpu_arch matches your GPU architecture. See for example this blog post to find your arch code.

Then build and run:

cd docker_cuspatial
./build.sh
./run.sh

You will find your HOME directory mounted to /host_home. If you require sudo to run docker, set environment variable SUDO like this

cd docker_cuspatial
SUDO=1 ./build.sh
SUDO=1 ./run.sh

If you have problems building you may use the uploaded dockerhub image by running

cd docker_cuspatial
./dockerhub_run.sh

This is not tested and you may run into issues with the environment variable gpu_arch. Possible workaround is to manually set it prior to compiling the code with build.sh.

Fix Python path

Add parsing folder to python path for correct imports.

source init_env.sh

Download Model Weights

Here you find the pre-trained models to reproduce the result from the paper.

• Cycle based simulated wireframe
• Polygon-HGT simulated wireframe
• Polygon-HGT joint prediction

You may for example put the model weights in the data folder, the rest of this README will assume you did.

Generate Annotations

To generate the annotations,

• Download Structured3D, see official website. You may use this script.
• Run python3 preprocessing/structured3D2planes_simple.py --help for instructions.

Build

To run the network, some C-code needs compiling.

./build.sh

Inference

There are a number of ways to run inference, see python3 scripts/test.py --help for details.

Run inference on test set with simulated wireframe

To run on the test set, do

cd scripts
python3 test.py \
--config-file ../config/Pred-plane-from-GT-GNN.yaml \
CHECKPOINT ../data/model_polygon_hgt_simulated.pth
OUTPUT_DIR ../runs/test

To run on the validation data, add the flag --val.

Run inference on your own images with joint prediction

To run on a set of images

cd scripts
python3 test.py \
--config-file ../config-files/Pred-simple-plane-S3D-GNN.yaml \
--img-folder <my-image-folder> \
CHECKPOINT ../data/model_polygon_hgt_joint.pth

and the result will be placed in <my-image-folder>/test.

Train

The different models are trained with their respective config file which exists in both simulated and joint prediction versions. The config files are placed in the folder config-files.

Train Polygon-HGT wit simulated wireframes

cd scripts
python3 train.py \
--config-file ../config-files/Pred-plane-from-GT-GNN.yaml

To monitor the training we have used W&B, but there is some support for tensorboard. W&B may be configured here.

Citation

If you use it in your research, please cite

@misc{gillsjö2023polygon,
      title={Polygon Detection for Room Layout Estimation using Heterogeneous Graphs and Wireframes},
      author={David Gillsjö and Gabrielle Flood and Kalle Åström},
      year={2023},
      eprint={2306.12203},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}