-
你好作者:你的提供ncnn模型很好,但是速度一般是1.1s左右;我有点得寸进尺,你能提供psnet文字检测的caffe模型吗。因为ncnn没加速,caffe是加速的,速度应该可以提高一倍多(cpu版本)
-
environment:Ubuntu 16.04
python 2.7.12
tensorflow 1.9.0
opencv 4.0
GPU:Tesla P100-PCIE-16GB *4
quesiton:
when i run the training,there is no result and no error,Always like this,why? ask for h…
-
为什么这个速度会比pse提升这么多呢,pse主要是在后处理部分耽误时间吗?模型大小好像跟pse差不多啊。但是pse后处理那部分使用c++写的,大概几百毫秒啊,为什么DB这个速度这么快?
-
Adam本身不就自带衰减效果吗,现在添加了MultiStepLR,貌似是屏蔽了Adam自带的衰减效果,然后设置几个阈值,达到一个阈值,lr = lr * gamma,可是这样有必要吗,为什么效果会优于Adam自身的衰减效果,查了好多资料也没搞明白个中关系,求大佬明示~
-
为啥 from .pse import pse_cpp, get_points, get_num在函数decode内部,局部导入,而不是全局导入呢,我在其他一个psenet项目中看到是全局导入,我对c++不是很了解
-
from .pse import pse_cpp
ImportError: Python version mismatch: module was compiled for Python 3.6, but the interpreter version is incompatible: 3.7.1 (default, Dec 18 2019, 23:13:02)
-
sas@sas-virtual-machine:~/桌面/tensorflow_PSENet-master$ python3 eval.py
DEBUG:matplotlib.pyplot:Loaded backend qt5agg version unknown.
DEBUG:matplotlib.pyplot:Loaded backend qt5agg version unknown.
…
-
https://github.com/liuheng92/tensorflow_PSENet/blob/5d057c6eccb79900b60fa349b4dddc756ec61052/pse/pse.cpp#L67-L79
-
你好,我想请教下这个代码在scut数据集上的效果怎么样,我用这个代码在icdar15数据集上测试过,效果能达到precision:0.8212,recall:0.7694,hmean:0.7944。
有一点不太明白的是论文中提出psenet可以痛过分割检测出任意形状的文字,但是我在看代码中生成mask的时候,是通过边框的四点坐标将原来的边框整个转换为了mask,是否对于scut数据集有不同的处理…
-
Hello -- are the weights for the pre-trained models the ones available on the struck out links in the readme -- that of ResNet50 and 152.
If not -- where could I download them?
Would be really …