RetinaNet

Tensorflow implement of RetinaNet

usage

details about code structure and how to use in usage.doc

pretrained weight

I use tensorflow's official weight, You can choose resnet weight or coco trained RetinaNet. Donwload them.

sorport

1.multigpu training
2.data augmentation(random crop/flip/rotate/color augment...)

spped

13fps/s in RTX2060.

performance

eval map(pretrained resnet50):~0.73 in voc07 (test) and trained in voc12(train+val)+voc07(train + val)
eval map(coco pretrained):~0.81 in voc07 (test) and trained in voc12(train+val)+voc07(train + val)
Batch size:4
data augmentation: flip

Results

referrence code

1.Tensorflow object API
2.Latter, I will put on them...