Deep3D

Real-Time end-to-end 2D-to-3D Video Conversion, based on deep learning.<br> Inspired by piiswrong/deep3d, we rebuild the network on pytorch and optimize it in time domain and faster inference speed. So, try it and enjoy your own 3D movies.

<div align="center"> <img src="./medias/wood_result_360p.gif"><br> </div>

Left is input video and right is output video with parallax.<br>

More examples:

• wood-1080p
• Journey to the West (86)

Inference speed

| Plan | 360p (FPS) | 720p (FPS) | 1080p (FPS) | 4k (FPS) | | :----------------------: | :--------: | :--------: | :---------: | :------: | | GPU (2080ti) | 84 | 87 | 77 | 26 | | CPU (Xeon Platinum 8260) | 27.7 | 14.1 | 7.2 | 2.0 |

Run Deep3D

Prerequisites

• Linux, Mac OS, Windows
• Python 3.7+
• ffmpeg 3.4.6+
• Pytorch 1.7.1
• CPU or NVIDIA GPU<br>

Dependencies

This code depends on opencv-python available via pip install.

pip install opencv-python

Clone this repo

git clone https://github.com/HypoX64/Deep3D
cd Deep3D

Get Pre-Trained Models

You can download pre_trained models from: [Google Drive] [百度云,提取码xxo0 ] <br> Note:

• 360p can get the best result.
• The published models are not inference optimized.
• Models are still under training, 1080p and 4k models will be uploaded in the future.

Run it!

python inference.py --model ./export/deep3d_v1.0_640x360_cuda.pt --video ./medias/wood.mp4 --out ./result/wood.mp4 --inv
# some video need to reverse left and right views (--inv)

Acknowledgements

This code borrows heavily from [deep3d] [DeepMosaics]