sunny2109 / SVLRM

The PyTorch implementation of CVPR2019 paper Spatially Variant Linear Representation Models for Joint Filtering

Spatially Variant Linear Representation Models for Joint Filtering

Dependencies

• Python = 3.8
• PyTorch = 1.5
• TensorBoard
• numpy
• os
• cv2
• PIL
• glob
• logging

Training

I trained and tested the model on a single NVIDIA RTX 2080Ti GPU. The training strategies are same as the paper.

• Command

#x4
python train.py --upscaling_factor 4
#x8
python train.py --upscaling_factor 8
#x16
python train.py --upscaling_factor 16

Testing

#x4
python test.py --upscaling_factor 4 --model weights/X4/model_192600_iter.pth
#x8
python test.py --upscaling_factor 8 --model weights/X8/model_177200_iter.pth
#x16
python test.py --upscaling_factor 16 --model weights/X16/model_192600_iter.pth

Results

• Quantitative results (RMSE)

Depth Image SR	SVLRM (paper)	Ours
x4	1.74	1.6722
x8	5.59	3.2172
x16	7.23	5.8599

Our all models results can be download from Baidu Cloud code: rjof or Google Drive

• Visual results (X8 depth sr)

on the left is output of the model, on the right is the corresponding ground truth image

img_001065 || RMSE:2.5730 || PSNR:39.9219 || SSIM:0.9825

img_001101 || RMSE:3.8905 || PSNR:36.3307 || SSIM:0.9748

img_001215 || RMSE:2.7525 || PSNR:39.3362 || SSIM:0.9839

img_001320 || RMSE:3.8696 || PSNR:36.3774 || SSIM:0.9751

img_001436 || RMSE:3.2991 || PSNR:37.7628 || SSIM:0.9763

Acknowledgements

• SVLRM_matlab
• SVLRM_Pytorch