the difference between SSD_tensorflow and your code

[hikapok]

the difference between SSD_tensorflow and your code

[HiKapok]

@liuqinying I have the same confustion at first. But rethink that why we need multi-layers prediction? We want different layer will detect different scale objects. So when we match anchors, we must ensure each anchor will be assigned the suitable scale object ground truth!

Consider one simple case when there is both a large dog and a small bottle in one image, if we match layer by layer then the bottle will be assigned to at least one anchor even at the most large scale feature map, so sub-optimal matching occured.

Hope that helps.