Follow this https://github.com/chuanqi305/MobileNet-SSD，mobilenet-ssd only get mAP 68% without MS-COCO pretraining。Even if the structure changes from ssd to ssdlite,why can it improve so much?
I trained mobilenet-ssdlite only on VOC with using imagenet pretrained model，it only can get 69.6%
thank you!
Follow this https://github.com/chuanqi305/MobileNet-SSD，mobilenet-ssd only get mAP 68% without MS-COCO pretraining。Even if the structure changes from ssd to ssdlite,why can it improve so much? I trained mobilenet-ssdlite only on VOC with using imagenet pretrained model，it only can get 69.6% thank you!