关于Size Regression标签的一些困惑

[murdockhou]

你好，感谢你的工作！关于Gaussian Kernels for Training， Size Regression部分，不是很明白里面的scalar s的作用，加上s后不是网络预测的 w_l, h_l, w_r, h_r值反而会更小更不利于网络预测吗？

[liuzili97]

Hi. Since the ttfnet will directly predict the box size on the original scale, the predictions may vary greatly (e.g., For the 512x512 input image, the predictions may vary from 10 to 200+), which may be harmful for training.

In order to avoid directly predicting large values, we introduce s here, as you said, to make the model predict smaller values. This helps to converge faster in the early stages of training. The s is set to 16, and it is not carefully selected.

[murdockhou]

@liuzili97 , thanks for your reply, i figured out, thanks.