Neural Scene Decoration

This is the code repository for the ECCV 2022 submission "Neural Scene Decoration from a Single Photograph".

Instructions

1. Download the Structured3D dataset and organize it in the following format:

Structured3D
	- data
		- scene_00000
		- scene_00001
		- ...
		- scene_03499
	- splits
		- bedroom_obj
		- living_room_obj

where splits are provided in this repository, and the scenes in data are downloaded from Structured3D.

2. Create virtual environment and install the dependencies. By default, the CUDA version used is 11.3.

conda create --name nsd python=3.8
pip install -r requirements.txt

3. Change dataset path (--data) in train_struct3d.sh to

4. Run train_struct3d.sh to start training.

Acknowledgement

This respository is modified from https://github.com/lucidrains/lightweight-gan.

Citation

Please cite our paper if you find it useful for your research:

@inproceedings{pang2022nsd,
  title={Neural Scene Decoration from a Single Photograph},
  author={Hong-Wing Pang and Yingshu Chen and Phuoc-Hieu Le and Binh-Son Hua and Duc Thanh Nguyen and Sai-Kit Yeung},
  year={2022},
  booktitle={ECCV},
}