search

zhreshold / mxnet-ssd

MXNet port of SSD: Single Shot MultiBox Object Detector. Reimplementation of https://github.com/weiliu89/caffe/tree/ssd
MIT License
764 stars 337 forks source link

confusion about MultiBoxTarget #65

Open xinghedyc opened 7 years ago

xinghedyc commented 7 years ago

Hi , I have a question about MultiBoxTarget layer. Why ssd uses MultiBoxTarget layer to do things like assigning positive and negative samples? I think these processings could be done when making a dataset, because all the default boxes are fixed, so I might just match these default boxes to image's ground truth bounding box. This way would save a lot of time during training. Is there anything wrong with this idea?

nopattern commented 7 years ago

refer to the paper about data augmentation.

zhreshold commented 7 years ago

Ground-truths changing all the time after augmentation. So it's better to put that into layers.

xinghedyc commented 7 years ago

@nopattern @zhreshold Thank you very much! I get it.