PeterVennerstrom
commented
2 years ago The gaussian kernel extends well beyond a 3*3 convolution or max pool kernel. With Beta of .54, the gaussian kernel extends across 54% of the width and height of the bounding box.
The benefit is increasing the number of i,j locations to supervise while assigning more of the i,j locations to object foreground. More supervised i,j locations over foreground makes the training more sample efficient.
In order to eliminate the need for the NMS, we adopt a similar strategy for center localization. Specifically, we further consider the aspect ratio of the box in the Gaussian kernel since the strategy that does not consider it in CenterNet is obviously sub-optimal.
为什么是centernet中不考虑长宽比的方法是sub-optimal的呢，能否给个原因，感觉ttfnet里面这样做意义不大，因为后面做max pool的卷积核也是3*3的，是正方形的