Wenhai Wang
Wenhai Wang
Yes, linear SRA will be used for all variants. We are making more improvements to linear SRA. All models will be released later. Thanks for your attention.
Thanks. I simply apply OHEM to the pixels in segmentation results. More detail can find in the discussion area of https://zhuanlan.zhihu.com/p/37884603, in which I disscus the same problem with fxwispig.
hi, the model (backbone: resnet50) only trained on icdar2015 can reach the F-measure 80.57. The model (backbone: resnet50) finetune from icdar2017 mlt can reach the F-measure 85.69. The long edge...
@tianzhuotao The output size of the model is (W / 4, H / 4), where W and H is the width and height of input image.
@liny23 I think the incomplete text line is a common case in natural scene images.
@duanjiaqi I have no result of this experiment. But I think it can make a little improvement in icdar2015.
The training set and val set for ICDAR 2017 MLT (http://rrc.cvc.uab.es/?ch=8).
We use multi scale training. The detail can be found in the new paper.
yes, just for testing.
Resnet50：P：73.7，R：68.2，F：70.8 Resnet152：P：75.3，R：69.2，F：72.2