Closed c-yn closed 1 year ago
Hello,
As mentioned in the repository of MIMO-UNet, most of the papers in image deblurring compare the asynchronous time, so we also measure in asynchronous time.
I compare Stripformer with SOTAs on the GoPro test set with a single RTX3090 GPU, Stripformer is still faster than these methods in synchronous time and achieves a good balance between Params, GFLOPs, and Time. The GFLOPs are measured on a 256x256 patch.
Methods | PSNR | Parms (M) | GFLOPs | Async-Time (ms) | Sync-Time (ms) |
---|---|---|---|---|---|
MPRNet (CVPR 2021) | 32.66 | 20 | 760 | 148 | 1410 |
MIMO-UNet++ (ICCV 2021) | 32.68 | 16 | 616 | 49 | 1115 |
Restormer (CVPR 2022) | 32.92 | 26 | 140 | 350 | 1116 |
Stripformer | 33.08 | 20 | 170 | 52 | 800 |
Very much thanks for your additional evaluation.
Dear author, I find that the inference speed reported in your paper is very fast. Do you think the testing time should be computed using torch.cuda.synchronize() as mentioned in the repository of MIMO-UNet? After adding this command, the speed decreases dramatically. Look forward to your opinion. Thanks.