Awesome

Indoor Scene Generation From a Collection of Semantic-Segmented Depth Images

This is the code for our paper Indoor Scene Generation From a Collection of Semantic-Segmented Depth Images.

Requirements

• cuda 10.1
• python 3.6.8
• Other requirements are list in requirements.txt

Data preparation

We provide the data processing scripts for Structured3D and Matterport3D dataset.

1. Download original dataset

Download Structured3D.

├──Structured3D
    ├── Structured3D_bbox.zip
    ├── Structured3D_0.zip
    ├── Structured3D_1.zip
    ├── ...
    ├── Structured3D_17.zip
    ├── Structured3D_perspective_0.zip
    ├── Structured3D_perspective_1.zip
    ├── ...
    ├── Structured3D_perspective_17.zip

Download Matterport3D.

├──Matterport3D/v1/scans
    ├── 1LXtFkjw3qL
    ├── 1pXnuDYAj8r
    ├── ...

2. Configure the dataset path

Configure data_dir and out_dir in examples/scenegen/tools/configs/DataGen-${dataset}-${type}.yaml, where dataset is Structured3D or Matterport3D, type is one of Bedroom, Living, Kitchen.

3. Generate the training data

Run

dataset=Structured3D;  # Structured3D or Matterport3D
type=Bedroom;  # Bedroom, Living or Kitchen

python process.py --task data_gen --cfg_path ./examples/scenegen/tools/configs/DataGen-${dataset}-${type}.yaml

The final training data train_image.records is stored in ${out_dir}/${label_type}/TrainViewData/.

Generative adversarial networks

1. Configure the experiment

Configure experiment in examples/scenegen/configs/Img2vol-${dataset}-${type}.yaml.

2. Train and test

Run

measure=Train;  # Train or Test
model_dir=./model_dir/${dataset}_${type}_model
python execute.py --example examples/scenegen --cfg_path ./examples/scenegen/configs/Img2vol-${dataset}-${type}.yaml --data_dir ${out_dir}/${label_type}/TrainViewData/ --model_dir ${model_dir} --log_dir ./log_dir/${dataset}_${type}_log --measure ${measure}

to train the model. Availabel arguments:

• --cfg_path: config file
• --data_dir: path of training data
• --model_dir: path to save trained model
• --log_dir: path to save training log
• --measure: Train or Test

Pre-trained models

The pre-trained models trained on Structured3D (bedroom, living room, kitchen) and Matterport3D (bedroom) dataset.

Generate scenes

Generate semantic scene volume

First set measure to Test, and then re-run execute.py to generate semantic scene volume from random noises. The generated scenes data eval_meta.npz is stored in ${model_dir}/eval/${training_epoch}.

Visualize the generated semantic scene volume

python process.py --task evaluation --cfg_path ./examples/scenegen/tools/configs/DataGen-${dataset}-${type}.yaml --eval_dir ${model_dir}/eval/${training_epoch}/ --output_dir ${model_dir}/eval/${training_epoch}/output/

The visualization is saved in ${model_dir}/eval/${training_epoch}/output/vis_3d.

Generate final indoor scenes

To generate the final 3D indoor scene by replacing each volumetric object instance in the volume with a CAD model retrieved from a 3D object database ShapeNet based on their type and volumetric shape.

First download ShapeNetCore v2 data from ShapeNet. Then run

shapenet_path=/PathToShapeNet/ShapeNetCore.v2.zip
python process.py --task retrieval --cfg_path ./examples/scenegen/tools/configs/DataGen-${dataset}-${type}.yaml --eval_dir ${model_dir}/eval/${training_epoch}/ --output_dir ${model_dir}/eval/${training_epoch}/output/ --shapenet_path ${shapenet_path}

The final generated scenes are saved in ${model_dir}/eval/${training_epoch}/output/retrieval.

Generated example

Citation

If you find our work useful for your research, please cite us using the bibtex below:

@article{yang2021indoor,
  title={Indoor Scene Generation from a Collection of Semantic-Segmented Depth Images},
  author={Yang, Ming-Jia and Guo, Yu-Xiao and Zhou, Bin and Tong, Xin},
  journal={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2021}
}