amdegroot / ssd.pytorch

A PyTorch Implementation of Single Shot MultiBox Detector
MIT License
5.14k stars 1.74k forks source link

Train from scratch #80

Open BunnyShan opened 6 years ago

BunnyShan commented 6 years ago

In the model perfomance comparison, the training from scratch means no imagenet pretrain?

DL-ljw commented 6 years ago

I am curious about it. If training from scratch can get such a good performance, why the researchers use the pre-trained model? Just for converging fast? I have always been thinking that training from scratch is not good as pretrained in mAP.

oarriaga commented 6 years ago

I believe the term "from scratch" means using the pre-trained-imagenet/modified VGG but starting with random weights the SSD-modules.

guantinglin commented 5 years ago

I think it means vgg base layers "pre-trained" on Imagenet and finetune the ssd extra layers with detection datasets(like VOC, COCO...)