wenston2006

Hard to repeat the training loss or accuracy as your trained model

3

Hi, thanks for your sharing of the codes at first. I tried to use some new data (20,000 pics) to train the 1_F model through finetuning. I got the test...

@tonghe90 请问训练时，文字识别的ground truth（文字的标签）在哪里输入？

2

如题，我在train.pt中没看到文字识别的ground truth，如果没有文字的标签，文字识别部分如何训练呢？

@tonghe90 输入的bbox大小有什么限制吗?

1

我把loss_4s和iou loss层都注释掉了,现在仅有文字识别的softmaxwithloss损失函数(mask loss和iou loss都不参与训练); 然后自己写了一个输入数据层,可以输出包含文字的图片(640*640大小), 作为gt的bbox的四个点的坐标以及文字的标签同时输出; 但是训练时候遇到segmentation fault, 提示内存越界; 请问输入给point bilinear layer的bbox大小有什么限制吗?64*8个采样点的条件下, 输入的bbox大小是否有什么要求?

@tonghe90 我这边没有12gb的显卡，可否用两块6gb的显卡（比如GTX1060）替代？

3

如题，用两块6gb（或8gb）的显卡和用一块12gb的显卡进行训练或测试有什么区别？

@tonghe90 train.pt中ignore_bbox是用来做什么的？

3

感觉在gt_bbox中把ignore_bbox去掉不就可以了吗？

@tonghe90 如何准备text/non-text 掩膜数据？

2

我的意思是 掩膜是单个字符的外接矩形框，还是要把字符沿着笔画边缘分割出来？

@tonghe90 训练数据的准备问题?

1

@tonghe90 感谢分享代码，我看了之前的issue #16 中，你提到训练数据包括(1) text/non-text region, (2) for every point in the text region, you need to calculate the distance between the current point to the four edges with an...

如题;我看了下网页上给出的测试结果,92%的字符准确率, 80%的字符串准确率; 相比有些算法(如CTC)的结果,似乎优势不明显

About some training and testing results of ABCNetV2

3

I tried to train ABCNetV2 to detect and recognize the chars on Chinese stamps. I modify the _configs/ReCTS/v2_chn_attn_R_50.yaml_ file, where I set: _MIN_SIZE_TRAIN: (128, 256, 384, 512, 640, 768) MAX_SIZE_TRAIN:...

@Yuliang-Liu Can you give some explanation about the setting of MIN/MAX_SIZE_TRAIN/TEST?

1

I found the default setting of MIN_SIZE_TRAIN in Base-BAText.yaml is (640, 672, ..., 896) . Can you give some explanation about this setting? I haven't found any information about this...