Yukang Wang
Yukang Wang
Sorry for replying late. These lines are added for robustness since the representatives may not quite close to each other for some large text instances. Here, `dict_p` also stores the...
Emmm... The main reason is that EAST uses a lightweight backbone network (PVANET). Note that with the same backbone network (VGG16), TextField runs at 6.0 fps on ic15, which is...
Maybe adding `boost::` to the `shared_ptr` could help.
About the environment: * gcc 5.4.0 * opencv 3.4.3 * opencv-python 3.4.3 Advice about training on total-text: * train `100k` and `40k` iterations for `lr=1e-4` and `lr=1e-5`, respectively, to reproduce...
Sorry for the late replay. I was too busy these days. It is a long story. In fact, the results in the paper are obtained with our original code, based...
Hi, thanks for your interest. From line 144 to 145 in `utils/criterion.py`: - `average pooling` is simply implemented by `sum` and `div`; - `unpooling` is then achieved by assigning the...
> 1. When computing the similarity between predictions and labels, any pre-operation should be used to process label data? For example, one-hot encoding or else? If you do, then where...
Same error with 2.3.6-cu114 when the number of non-empty voxels is very large. And sampling on inputs works for me.
Please leave your email address and we will send it to you.
I think any label tools for image segmentation, e.g. labelme, can do it.