search

ZJULearning / ttfnet

Apache License 2.0
485 stars 99 forks source link

question about loss #10

Closed chaos1992 closed 5 years ago

chaos1992 commented 5 years ago

In this paper, loss is composed of hm loss and wh loss, why did you delete reg loss used in CenterNet?

chaos1992 commented 5 years ago

If I add reg loss in ttfnet, can I get better performance?

liuzili97 commented 5 years ago

In CenterNet, the offset of the center will cause the offset of the final prediction box. So one needs to add the offset to get better performance. In TTFNet, instead of predicting the width and height of the object, we predict the distance from the center to the four sides of the box. So this is equivalent to integrating the offset into the regression. Adding offset is not needed.

chaos1992 commented 5 years ago

thanks for you reply