about pvanet training about east HOT 16 CLOSED

I have tried pva, it can get comparable result with resnet(need pretrain on imagenet).
IOU is just part of the loss, you should also see the regression part, during training you should set a balance parameter between the two losses. I do not recommend rotation augmentation as it may cause some un-agreement of text orientation and the start point of text polygon.

from east.

thanks! so you just use random scaling to augment data as training with res_net-50?
In my training I use Iou loss and cosine loss to regress rbox and class balancing softmax loss to predict the mask of shrinked text polygon, what is the regression part? whether you mean that in pva training you used other rbox regression loss besides Iou and cosine loss?

from east.

Do you use imagenet pretrained models?

from east.

yes I do

from east.

do you have a code repo? maybe some bugs..

from east.

you can check this code first, its easy to run

from east.

OK I go to check.. 5555 so sad

from east.

sorry to bother you again, will dice loss influence much in pva training? I implemented entirely base upon paper's method

from east.

In my experiment, dice loss perform much better..

from east.

@argma Hi，what do you mean by

some un-agreement of text orientation and the start point of text polygon.

As far as I know, the rotation augmentation is used in many recent papers on text detection. Is there any difference between EAST and the other algorithms on dealing with the rotation augmentation?

from east.

@zxDeepDiver , in east, there is upper, right, down and left side of a polygon, when rotated, this order can change, i think other methods encounter this problem too.

from east.

hi! argman, now my iou average training loss over one mini-batch can reduce to 0.18( the loss I used is 1.0 - Iou, 0,18 means that training Iou is 1 - 0.18) and my validation average Iou loss over all test sets which has ignored 'do not care' region reduces to 0.44.
could you share your training and validation loss in the last few iterations?

from east.

I do not test that, why do you test the final output ?

from east.

@argman do you have a tensorflow slim code for pvanet ?

from east.

do you have a tensorflow slim implementation code for pvanet?

from east.

I have tried pva, it can get comparable result with resnet(need pretrain on imagenet).

Could you please advice on how you did the pre training? More specifically:

• How to isolate the backbone and train only this
• Any suggested subset of Imagenet for training

Thanks.

from east.