search

kaix90 / DCTNet

457 stars 97 forks source link

What's the difference between DCT in 8x8 patch and a parameter defined 8x8 convolution kernel? #2

Closed maxenceliu closed 4 years ago

maxenceliu commented 4 years ago

totally equals k=8, s=8, parameter defined layer

kaix90 commented 4 years ago

You are correct from the computation perspective. But the result of DCT is more interpretable since different frequency components are separated. While the result of convolution is less interpretable since different frequency components are mixed.

maxenceliu commented 4 years ago

You are correct from the computation perspective. But the result of DCT is more interpretable since different frequency components are separated. While the result of convolution is less interpretable since different frequency components are mixed.

Interpretable is not a objective word. Especially for the first level of CNN. Many papers have interpret the function of low level convolution.

DCT is definitely a certain form of convolution. DCT = ONLY ONE LAYER CONVOLUTION. This paper just use another way to explain convolution, because DCT are just kernels with orthogonality. The conclusion will mislead readers. This kind of decoration and ｌacking depth discussion will become a bluff in DL research.

If this paper could deeply discuss the effectiveness of manual designed low level kernels (Such as DCT,　Haar-like, HOG, LBP...) with different parameters, it will be valuable.

kaix90 commented 4 years ago

You are correct that DCT is definitely a certain form of convolution. But our paper is not aiming to use another way to explain convolution. Instead, we explore the importance of each frequency component in several tasks. DCT is just a tool to transform data to the frequency domain. So please point out which conclusion is misleading readers, then we can discuss further. Thanks.

On Mar 11, 2020, at 6:39 PM, maxenceliu notifications@github.com wrote:

 You are correct from the computation perspective. But the result of DCT is more interpretable since different frequency components are separated. While the result of convolution is less interpretable since different frequency components are mixed.

DCT is definitely a certain form of convolution. This paper just use another way to explain convolution, because DCT are just kernels with orthogonality. The conclusion will mislead readers. This kind of decoration and ｌacking depth discussion will become a bluff in DL research.

If this paper could deeply discuss the effectiveness of manual designed low level kernels (Such as DCT,　Haar-like, HOG, LBP...) with different parameters, it will be valuable.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

maxenceliu commented 4 years ago

You are correct that DCT is definitely a certain form of convolution. But our paper is not aiming to use another way to explain convolution. Instead, we explore the importance of each frequency component in several tasks. DCT is just a tool to transform data to the frequency domain. So please point out which conclusion is misleading readers, then we can discuss further. Thanks.

explore the importance of each frequency component is more accurate than Learning in the Frequency Domain. But still needs further discussion.

kaix90 commented 4 years ago

Thanks for your suggestion.