Is it possible to get good results without pretraining?

[sherif7810]

I over-fit on a small sample of TotalText, until I got a loss of about 0.1 in 6 images, but the demo file with the trained checkpoint doesn't show any contours.

[sherif7810]

Reducing thresholds adds red dots to output with the runtime of demo increased a lot.

[princewang1994]

It seems that model is underfitting but not overfitting. The loss should be smaller than 0.1. For small dataset overfit, you can try smaller learning rate and training more epochs.

[sherif7810]

Do you know a configuration (Number of samples, batch size, epochs, learning rate, etc) that is guaranteed to overfit successfully?
I can't increase batch size to 4.

[sherif7810]

Re-downloading the repository fixed the issue. It seems I broke something.

[sherif7810]

You can close the issue. Thanks for help.

[dongzhi0312]

为什么输出的识别结果图片好像被重复了一次

[princewang1994]

@dongzhi0312 这个地方代码中有点小问题，在demo.py里面这个部分，这了predict和gt使用了相同的coutour变量，其实下面那张图片应该是使用meta变量中的坐标点的，你发的这张图片的上下两个应该都是predict，我会在接下来的版本中改正~

[dongzhi0312]

把这段代码改写成下面这样，解决了图像重复一次的问题
demo.py 中的85 86 行注释，88行改为 cv2.imwrite(path, predict_vis) 解决了图像重复的问题。图像重复问题由 86的 np.concatenate() 造成的。

[sherif7810]

What is the best training configuration for generalization on test set? This is with batch size of 2:

Blue: Training. Red: Testing.