NonCuboidRoom

Paper

Learning to Reconstruct 3D Non-Cuboid Room Layout from a Single RGB Image

Cheng Yang*, Jia Zheng*, Xili Dai, Rui Tang, Yi Ma, Xiaojun Yuan.

[Preprint] [Supplementary Material]

(*: Equal contribution)

Installation

The code is tested with Ubuntu 16.04, PyTorch v1.5, CUDA 10.1 and cuDNN v7.6.

# create conda env
conda create -n layout python=3.6
# activate conda env
conda activate layout
# install pytorch
conda install pytorch==1.5.0 torchvision==0.6.0 cudatoolkit=10.1 -c pytorch
# install dependencies
pip install -r requirements.txt

Data Preparation

Structured3D Dataset

Please download Structured3D dataset and our processed 2D line annotations. The directory structure should look like:

data
└── Structured3D
    │── Structured3D
    │   ├── scene_00000
    │   ├── scene_00001
    │   ├── scene_00002
    │   └── ...
    └── line_annotations.json

SUN RGB-D Dataset

Please download SUN RGB-D dataset, our processed 2D line annotation for SUN RGB-D dataset, and layout annotations of NYUv2 303 dataset. The directory structure should look like:

data
└── SUNRGBD
    │── SUNRGBD
    │    ├── kv1
    │    ├── kv2
    │    ├── realsense
    │    └── xtion
    │── sunrgbd_train.json      // our extracted 2D line annotations of SUN RGB-D train set
    │── sunrgbd_test.json       // our extracted 2D line annotations of SUN RGB-D test set
    └── nyu303_layout_test.npz  // 2D ground truth layout annotations provided by NYUv2 303 dataset

Pre-trained Models

You can download our pre-trained models here:

• The model trained on Structured3D dataset.
• The model trained on SUN RGB-D dataset and NYUv2 303 dataset.

Structured3D Dataset

To train the model on the Structured3D dataset, run this command:

python train.py --model_name s3d --data Structured3D

To evaluate the model on the Structured3D dataset, run this command:

python test.py --pretrained DIR --data Structured3D

NYUv2 303 Dataset

To train the model on the SUN RGB-D dataset and NYUv2 303 dataset, run this command:

# first fine-tune the model on the SUN RGB-D dataset
python train.py --model_name sunrgbd --data SUNRGBD --pretrained Structure3D_DIR --split all --lr_step []
# Then fine-tune the model on the NYUv2 subset
python train.py --model_name nyu --data SUNRGBD --pretrained SUNRGBD_DIR --split nyu --lr_step [] --epochs 10

To evaluate the model on the NYUv2 303 dataset, run this command:

python test.py --pretrained DIR --data NYU303

Inference on the customized data

To predict the results of customized images, run this command:

python test.py --pretrained DIR --data CUSTOM

Citation

@article{NonCuboidRoom,
  title   = {Learning to Reconstruct 3D Non-Cuboid Room Layout from a Single RGB Image},
  author  = {Cheng Yang and
             Jia Zheng and
             Xili Dai and
             Rui Tang and
             Yi Ma and
             Xiaojun Yuan},
  journal = {CoRR},
  volume  = {abs/2104.07986},
  year    = {2021}
}

LICENSE

The code is released under the MIT license. Portions of the code are borrowed from HRNet-Object-Detection and CenterNet.

Acknowledgements

We would like to thank Lei Jin for providing us the code for parsing the layout annotations in SUN RGB-D dataset.