search

QinSY123 / 2024-MambaVC

Code for MambaVC: Learned Visual Compression with Selective State Spaces
32 stars 1 forks source link

RD-data #1

Open Nanxia001 opened 4 months ago

Nanxia001 commented 4 months ago

请问是否方便提供下MambaVC在Kodak测试集上RD_Curve各点的具体数值？以及什么时候可以分享训好的模型进行测试？

QinSY123 commented 4 months ago

我们正在MLIC对应更高质量数据集上重新训练，训练好后我们会放出模型和数据。

faymek commented 3 months ago

Not consistent result with MLIC+

test

MambaVC 
X   Y
0.891213389 38.02570573
0.631520223 36.07381748
0.440167364 34.20996183
0.30446304  32.54554182
0.19902371  31.01809218

MLIC+   
X   Y
0.811157601 37.3421167
0.591492329 35.63899971
0.427475593 33.96167494
0.282008368 32.16064783
0.182426778 30.64570508
QinSY123 commented 3 months ago

Not consistent result with MLIC+

test

MambaVC   
X Y
0.891213389   38.02570573
0.631520223   36.07381748
0.440167364   34.20996183
0.30446304    32.54554182
0.19902371    31.01809218

MLIC+ 
X Y
0.811157601   37.3421167
0.591492329   35.63899971
0.427475593   33.96167494
0.282008368   32.16064783
0.182426778   30.64570508

The MLIC+ data in the article was also trained on the Flickr30k dataset, whereas the original MLIC+ paper was trained on a dataset of 100,000 high-resolution images. Therefore, the numerical values we obtained are inconsistent with those reported in the original paper. We have already communicated this with the MLIC+ authors, and we are currently retraining MambaVC on the high-resolution dataset.

galijipaifan commented 3 months ago

你好，我注意到论文中（mixedcvpr23）在kodak数据集上的数据与原论文中的结果不一致，原论文在bpp0.6时，psnr大概在35.9db而您论文中的数据似乎不到35.6db。

QinSY123 commented 3 months ago

你好，我注意到论文中（mixedcvpr23）在kodak数据集上的数据与原论文中的结果不一致，原论文在bpp0.6时，psnr大概在35.9db而您论文中的数据似乎不到35.6db。

TCM原论文给出的数据曲线是Large的结果，而我们使用small在Flickr上进行了重新训练。由于之前的不同方法是在不同数据集上训练的，我们需要在一个统一的数据集上重新训练消除不同训练集带来的影响，TCM原论文训练集有三十万张图片，而Flickr30k训练集只有三万张图片