xiaxiaoling

in FOTS.PyTorch / FOTS / data_loader / dataset.py If crop the picture, rectangles is not defined in code. And if not crop the picture, rectangles in line 179 is already...

recognition will be very slow when boxes are many

4

when i use a pic with 284 text, the roi and recognition cost lost of time, and if the boxes more than 300, because the reconigion without batchsize, the code...

icdar2015上训练效果

16

使用icdar2015数据来训练，参数使用作者默认的，为什么loss都很高呢，只训练256大小的 Epoch 1/24 1125/1125 [==============================] - 131s 117ms/step - loss: 1.0463 - val_loss: 1.1207 Epoch 00001: val_loss improved from inf to 1.12069, saving model to model/weights_3T256.001-1.121.h5 Epoch 2/24 1125/1125 [==============================]...

在处理训练数据的时候label存在问题

12

如果使用的数据label长度都相同的时候不会有问题，换了新的数据，label长度不同，因为新的batch是直接把新的值覆盖在上一个batch的label上的，train.py ：104行，就会造成不符合事实的label。 应该加一句 labels[i, len(str):] = 10000 在105行，还原padding的值

Why the hmean is lower than EAST

1

https://github.com/argman/EAST train in EAST the hmean is 0.80 the precision is 0.846 the recall is 0.77, the detection part of FOTS is vary similar to EAST, and the result of...

How to prepare training data for retraining?

7

I use the icdar data , i have changed it into : 0 0.5203125 0.21458333333333332 0.5325 0.135 0 PROPER 0 0.5234375 0.5104166666666666 0.535 0.205 0 FOOD 0 0.521875 0.7947916666666667 0.53...

cpp的pse有些问题

3

某些图片在cpp的pse过程会卡住，没报错一直卡住，改成pypse就可以正常运行。pypse代码里有个bug，25行应该改成for i in range(label_values): 之前是for i in label_values:，label_values是个int