search

yun-liu / RCF

Richer Convolutional Features for Edge Detection
Other
753 stars 259 forks source link

I wonder why each conv layer is connected to a conv layer with channel depth 21. #7

Closed MaFuyan closed 6 years ago

MaFuyan commented 6 years ago

Why 21?

MaFuyan commented 6 years ago

thx ：）

yun-liu commented 6 years ago

In fact, there isn't any special reason for using channel depth 21. I will provide a new version by replacing 21 with another number to make it reasonable.

MaFuyan commented 6 years ago

@yun-liu Thanks for your reply. Actually, I have been thinking this problem. I want to figure out the realationship between the depth and the final result. Can you tell me something about it?

yun-liu commented 6 years ago

I find there is little effect on the results when the number of channels is in some range, such as tens of channels.

MaFuyan commented 6 years ago

@yun-liu Grateful. I will adopt this net to another task. If I encouter other problems, I will contact you in time. Thanks.

MaFuyan commented 6 years ago

@yun-liu Your work also benefits a lot from dilated convolutions, right? Why not remove the last max pooling layer (pool4)?

yun-liu commented 6 years ago

There is not so much benefit of the dilated convolutions for edges, and I have tested it for HED. We shouldn't remove pool4. If we remove it, the pre-trained ImageNet model will be not suitable for conv5 because the representation of the network has been changed.