Will resnet50 or pvanet works better? about pixel_link HOT 13 OPEN

@aravinthmuthu i try to use modified VGG Net as backbone, after cut a lot of parameters, the final mAP is still at 82.5%, PixelLink network is a little bit tricky.

from pixel_link.

@baby313
I got worse result too with resnet50. Which endpoints did you use in resnet?

from pixel_link.

Pixel_link learns very minimalistic representations with limited receptive fields. Hence ResNet might be an overkill. Has anyone experimented with alexNet as the base model?

from pixel_link.

@BowieHsu
PixelLink is really tricky, never seen this kind of network before. My final mAP is 83.3 with only icdar15 dataset. I tried to raise this score by inserting batch norm or change hyper parameters, but it never succeeded yet so far. I'll share it if there is any improvement.

from pixel_link.

@baby313 @GodOfSmallThings @aravinthmuthu @BowieHsu @comzyh
Did you use the pretained resnet on Imagenet? and could you share the best score based on resnet?

from pixel_link.

@lizzyYL
In my case, resnet50 performed worse(~7x%) with paper's training setting.
And since learning rate is relatively so high(1e-2), I'm not sure ImageNet pretrained model would work.

from pixel_link.

@GodOfSmallThings Thank you for your reply.
My result is 68%-70%(80000 iterations), and the resnet did not been pretrained.

from pixel_link.

~7x% @ not pretrained @lizzyYL

from pixel_link.

@GodOfSmallThings @BowieHsu @lizzyYL
i can't evaluate my result
please help me to calculate precision and recall on my result

from pixel_link.

@mn7142 You can use the test scripts to build an offline test platform，
or use the online testing : http://rrc.cvc.uab.es/?ch=4&com=evaluation&task=1&gtv=1

from pixel_link.

Replacing vgg with resnet-50, why did it deteriorate? Who can explain to me? thanks .

from pixel_link.

thank you @lizzyYL
how to evaluate on other dataset?

from pixel_link.

Have you add dilation to the second last conv?

from pixel_link.