Awesome

Diverse Plausible 360-Degree Image Outpainting for Efficient 3DCG Background Creation

Diverse Plausible 360-Degree Image Outpainting for Efficient 3DCG Background Creation (CVPR 2022)<br/> Naofumi Akimoto, Yuhi Matsuo, Yoshimitsu Aoki<br/>

arXiv | BibTeX | Project Page | Supp Video

Requirements

A suitable conda environment named omnidreamer can be created and activated with:

conda env create -f environment.yaml
conda activate omnidreamer

Trained models

Please send us an email. We will send you the URL for downloading. You may distribute the trained models to others, but please do not reveal the URL.

Running trained models

• Put trained weights under logs/
• Comment out ckpt_path to VQGAN models from each {*}-project.yaml

Inference examples

Sampling for a comparison against 360IC

CUDA_VISIBLE_DEVICES=0 python sampling.py \
--config_path logs/2021-07-27T05-57-41_sun360_basic_transformer/configs/2021-07-27T05-57-41-project.yaml \
--ckpt_path logs/2021-07-27T05-57-41_sun360_basic_transformer/checkpoints/last.ckpt \
--config_path_2 logs/2021-07-27T10-49-57_sun360_refine_net/configs/2021-07-27T10-49-57-project.yaml \
--ckpt_path_2 logs/2021-07-27T10-49-57_sun360_refine_net/checkpoints/last.ckpt \
--outdir outputs/test

Sampling for a comparison against SIG-SS

CUDA_VISIBLE_DEVICES=0 python sampling.py \
--config_path logs/2021-07-27T05-57-41_sun360_basic_transformer/configs/2021-07-27T05-57-41-project.yaml \
--ckpt_path logs/2021-07-27T05-57-41_sun360_basic_transformer/checkpoints/last.ckpt \
--config_path_2 logs/2021-07-27T10-49-57_sun360_refine_net/configs/2021-07-27T10-49-57-project.yaml \
--ckpt_path_2 logs/2021-07-27T10-49-57_sun360_refine_net/checkpoints/last.ckpt \
--mask_path assets/90binarymask.png \
--outdir outputs/test

Sampling for a comparison against EnvMapNet

CUDA_VISIBLE_DEVICES=0 python sampling.py \
--config_path logs/2021-08-12T03-27-04_sun360_basic_transformer/configs/2021-08-12T03-27-04-project.yaml \
--ckpt_path logs/2021-08-12T03-27-04_sun360_basic_transformer/checkpoints/last.ckpt \
--config_path_2 logs/2021-08-12T03-42-53_sun360_refine_net/configs/2021-08-12T03-42-53-project.yaml \
--ckpt_path_2 logs/2021-08-12T03-42-53_sun360_refine_net/checkpoints/last.ckpt \
--mask_path assets/90binarymask.png \
--outdir outputs/test

Training

We train the four networks separately (VQGAN_1, VQGAN_2, Transformer, and AdjustmentNet).
Order of training

• For the training of the Transformer, trained VQGAN_1 and VQGAN_2 are required.
• For the training of AdjustmentNet (RefineNet), trained VQGAN_2 is required.
• Therefore, VQGAN_1 and VQGAN_2 can be trained in parallel, and Transformer and AdjustmentNet (RefineNet) can also be trained in parallel.

Sample Commands

• VQGAN_1
  See configs/sun360_comp_vqgan.yaml for the details. Run it with image_key: concat_input, concat_input: True, in_channels: 7, and out_ch: 7. We trained 30 epochs totaly (only for transformer, 15).

CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0,1 python main.py --base configs/sun360_comp_vqgan.yaml -t True --gpus 0,1

• VQGAN_2
  Use configs/sun360_comp_vqgan.yaml for VQGAN_2, that is the same configuration file, but set image_key: image, concat_input: False, in_channels: 3, and out_ch: 3.

CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0,1 python main.py --base configs/sun360_comp_vqgan.yaml -t True --gpus 0,1

• Transformer
  Write the paths of the trained VQGAN_1 and VQGAN_2 in the yaml configuration file configs/sun360_basic_transformer.yaml. Even if you train VQGAN with 256x256 images before, you can train a transformer with 256x512. If using 256x512, set block size: 1024.

CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0,1 python main.py --base configs/sun360_basic_transformer.yaml -t True --gpus 0,1

• RefineNet(AdjustmentNet)
  Write the paths of the trained VQGAN_2 in the yaml configuration file configs/sun360_refine_net.yaml.

CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=0,1 python main.py --base configs/sun360_refine_net.yaml -t True --gpus 0,1

Development environment

• Ubuntu 18.04
• Titan RTX or RTX 3090
• CUDA11

License

This repo is built on top of VQGAN. See the license.

BibTeX

@inproceedings{akimoto2022diverse,
    author    = {Akimoto, Naofumi and Matsuo, Yuhi and Aoki, Yoshimitsu},
    title     = {Diverse Plausible 360-Degree Image Outpainting for Efficient 3DCG Background Creation},
    booktitle   = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year      = {2022},
}