Skip to content
forked from NJUVISION/AWnet

A Dual Camera System for High Spatiotemporal Resolution Video Acquisition (TPAMI 2020)

License

Notifications You must be signed in to change notification settings

chengming2/AWnet

 
 

Repository files navigation

AWnet

A Dual Camera System for High Spatiotemporal Resolution Video Acquisition

Project | Paper | video

Ming Cheng, Zhan Ma, M. Salman Asif, Yiling Xu, Haojie Liu, Wenbo Bao, and Jun Sun

IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)

image

Installation

The code has been tested with Python 3.7, PyTorch 1.0, CUDA 10.1 and Cudnn 7.6.4.

Once your environment is set up and activated, generate the Correlation package required by PWCNet:

$ cd correlation_package_pytorch1_0
$ sh build.sh

Training

Training data

Vimeo90K REDS

data process

Downsample the images with bicubic downsampling method by 4 times (or 8 times). Then, upsample the down-scaled images by 4 times (or 8 times), which will be used as the low-quality input of AWnet. The adjacent original high-quality frames can be used as the reference frames.

Training steps

STEP 0: Pre-load FlowNet

wget http://vllab1.ucmerced.edu/~wenbobao/DAIN/pwc_net.pth.tar

STEP 1: Finetune FlowNet
STEP 2: Pre-train FusionNet
STEP 3: End-to-end Finetune FlowNet and FusionNet

Demos

Image demos

These images are captured with our dual iPhone 7 cameras.

Video Demos

Video Demos (page 17)

Different illumination conditions: High Light Illumination | Medium Light Illumination | Low Light Illumination

Single-Reference vs Multi-Reference: Simulated data | Real data

Unfortunately, I lost the models of one-reference AWnet (AWnet_1). Luckily, I saved the models of two-reference AWnet (AWnet_2), as shown below: Model without noise | Model with noise (0.008)

About

A Dual Camera System for High Spatiotemporal Resolution Video Acquisition (TPAMI 2020)

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 65.8%
  • Cuda 24.4%
  • C++ 7.6%
  • Shell 2.2%