a question.

[wsp11]

the bpp you showed is tex_bpp.
But I think the bpp should be tex_bpp + feat_bpp.
I wanna know your opinion.

[ceynri]

The assumption here is that the feature information is necessary (for deep learning tasks downstream), on this premise to discuss the storage of the original image information.
Therefore, no matter only texture information or original image information, it is necessary to store an additional feature information.
The comparison of bpp can be understood as "tex_bpp + feat_bpp" and "original_bpp + feat_bpp", so the feat_bpp was omitted.

[wsp11]

但是这与论文中所描述的就不一样了吧，论文中特征信息也是要通过量化和熵编码然后进行传输，然后图片纹理部分用更少的bpp，使得整体图片获得更好的效果。

[ceynri]

可能是论文有水分、或者我们的算法实现没能很好的复现论文的效果吧，如果要特征数据+纹理数据去和传统算法做对比，我们的实验得到的结果是差不多的（我记得输出出来的特征数据占的存储空间就不小了，留给纹理数据的空间不多）