Skip to content

Latest commit

 

History

History
88 lines (64 loc) · 4.7 KB

README.md

File metadata and controls

88 lines (64 loc) · 4.7 KB
Raw

News

  • 2023/7/31 We release the complete training code for our method on the COCO dataset!
  • 2022/8/4 We release the code and the pretrained model for AdaMPI. Use it to create 3D photo from single in-the-wild 2D images!
  • 2022/7/22 We release the code and the pretrained inpainting model for the Warp-Back strategy.

AdaMPI

This is a PyTorch implementation of the following paper:

Single-View View Synthesis in the Wild with Learned Adaptive Multiplane Images, SIGGRAPH 2022.

Yuxuan Han, Ruicheng Wang and Jiaolong Yang

Project Page | Video | Paper

Abstract: This paper deals with the challenging task of synthesizing novel views for in-the-wild photographs. Existing methods have shown promising results leveraging monocular depth estimation and color inpainting with layered depth representations. However, these methods still have limited capability to handle scenes with complex 3D geometry. We propose a new method based on the multiplane image (MPI) representation. To accommodate diverse scene layouts in the wild and tackle the difficulty in producing high-dimensional MPI contents, we design a network structure that consists of two novel modules, one for plane depth adjustment and another for depth-aware color prediction. The former adjusts the initial plane positions using the RGBD context feature and an attention mechanism. Given adjusted depth values, the latter predicts the color and density for each plane separately with proper inter-plane interactions achieved via a feature masking strategy. To train our method, we construct large-scale stereo training data using only unconstrained single-view image collections by a simple yet effective warp-back strategy.

Environment

run with docker:

IMG=mpi dev

create conda env

conda env create -f environment.yml
conda activate mpi

run generation video in cpu001

python gen_3dphoto_video.py --video_path debug/yard_friend.mp4 --output_path /mnt/gz01/experiment/liuye/mpi
python gen_3dphoto_rgbd_rs.py --session_path data/20231115T033013Z --output_path data/20231115T033013Z_mpi
python gen_3dphoto_rgbd_iphone.py --session_path data/2023_11_15_14_31_58 --output_path data/2023_11_15_14_31_58_mpi

Run in Determined

VIDEO_PATH=/mnt/data/yeliu/MPI/forest.mp4
OUTPUT_PATH=/mnt/data/yeliu/MPI/forest
WEIGHT_PATH=/mnt/data/yeliu/models/mpi_weight

python gen_3dphoto_video.py \
--video_path ${VIDEO_PATH} --output_path ${OUTPUT_PATH} \
--ckpt_path ${WEIGHT_PATH}/adampi_32p.pth --cache_dir ${WEIGHT_PATH} --device cuda

python make_video.py --output_path /mnt/data/yeliu/MPI/war

generate VR LR video

python gen_vr_180.py --output_path /mnt/gz01/experiment/liuye/mpi/video_mpi_yangtuo
python gen_vr_180.py --output_path /mnt/ml-experiment-data/yeliu/MPI/ipman --video_path /mnt/ml-experiment-data/yeliu/MPI/ipman_audio.mp4
python gen_vr_180.py --output_path /mnt/data/yeliu/MPI/forest --video_path /mnt/data/yeliu/MPI/forest.mp4

Document

This repo contain two parts:

  • The train and inference code for the AdaMPI. You can use it to generate 3D photo from single-view in-the-wild image input, as shown in the teaser gif. See AdaMPI.md for more details.

  • The clean and abstract implementation of the Warp-Back strategy, you can use it to generate multi-view training data to train your own view synthesis model (e.g. SynSin, WorldSheet) using only single-view image collections like COCO dataset. See Warp-Back.md for more details.

Contact

If you have any questions, please contact Yuxuan Han ([email protected]).

License and Citation

This repository can only be used for personal/research/non-commercial purposes. Please cite the following paper if this model helps your research:

@inproceedings{han2022single,
    author = {Han, Yuxuan and Wang, Ruicheng and Yang, Jiaolong},
    title = {Single-View View Synthesis in the Wild with Learned Adaptive Multiplane Images},
    booktitle = {ACM SIGGRAPH},
    year={2022}
}

Acknowledgments

  • The code for inference and training is heavily borrowed from MINE, we thank the author for their great effort.
  • We thank Varun Jampani and Meng-Li Shih for helpful suggestions and discussions to evaluate their method.