Implementation of the Multiple Attention Swin Transformer Enhanced Residual GAN in PyTorch on DemDataset. This repository hosts the training and validation code for super-resolution of Depth Elevation Maps using MASTER GAN.
The introduced novel model uses multiple attention modules together, along with Residual skip connections and a simple CNN based feature extractor to create high resolution depth elevation map from a given low resolution depth elevation map. The model architecture is shown in:
The introduced model achieves state-of-the-art results across all metrics on the test set of DemDataset. The table below shows the performance of various models discussed, along with existing state-of-the-art results on the dataset.
| Model | PSNR | SSIM |
|---|---|---|
| Swin IR | 30.753 | 0.907 |
| SRResNet | 26.876 | 0.845 |
| D-SRGAN | 30.828 | 0.9172 |
| ESPCN | 29.078 | 0.877 |
| SRCNN | 30.630 | 0.907 |
| MATSRGAN | 31.024 | 0.908 |
-
Author 1:
- Name: Azhan Mohammed
- LinkedIn: https://www.linkedin.com/in/azhanmohammed/
- Personal Website: https://www.azhanmohammed.xyz
-
Author 2:
- Name:
- LinkedIn:
- Personal Website: