makbari7 / DSSLIC

DSSLIC: Deep Semantic Segmentation-based Layered Image Compression
45 stars 19 forks source link

Cannot get the result in the paper with the pretrained model #2

Open qjfytz opened 5 years ago

qjfytz commented 5 years ago

I have tried to run the test command as shown in the home page: python test.py --name Cityscapes_model --dataroot ./datasets/cityscapes/ --label_nc 35 --loadSize 1024 --resize_or_crop scale_width --batchSize 1 --how_many 50

The pretrained models and datasets are downloaded from the link offered in the readme files. After get the generated images, I have tested the average PSNR and MS-SSIM of all images, which are MS-SSIM 0.9587, and PSNR 29.36. MS-SSIM values are calculated using https://github.com/jorge-pessoa/pytorch-msssim/blob/master/pytorch_msssim/__init__.py, PSNR values are calculated with own script.

These are far worse than the result presented in the paper, where PSNR and MS-SSIM can be 34dB and 0.98 with 0.1 bpp , as shown in Fig.3. So I am wondering if there is any problem in my test. Could you please figure it out?

And I cannot find a script to get the compressed files, so that I cannot calculated the bpp. So could you please tell me how to evaluate the compression ratio of the model?

ChenWei007HB commented 4 years ago

Hi, qjfytz, I met the same issue as you, when tested on the Kodak dataset, the result can not surpass the BPG method on PSNR and MSSSIM, I wonder whether your problem have been solved? If solved, how can you work it out?