RetinaNet support

[miliadis]

This is a PR to support RetinaNet, a state-of-the-art one-stage object detector. The model (ResNet 50 with 800 pixels) achieved AP of 35.3 instead of 35.7 reported on the paper.

1. you can download my PyTorch weights here (AP 35.3)
2. this code allows you also to load Caffe2's weights of RetinaNet. You can download the weights here (AP 35.7)

Details of results are shown below:

RetinaNet

retinanet-R-50-FPN_1x

• Testing command with PyTorch weights:

  python tools/test_net.py \
  --dataset coco2017 --cfg configs/baselines/retinanet_R-50-FPN_1x.yaml \
  --load_ckpt model_step179999.pth

• Testing command with Detectron weights:

  python tools/test_net.py \
  --dataset coco2017 --cfg configs/baselines/retinanet_R-50-FPN_1x.yaml \
  --load_detectron retinanet_pretrained.pkl

Box
source	AP50:95	AP50	AP75	APs	APm	APl	AR1	AR10	AR100	ARs	ARm	ARl
PyTorch	35.3	54.6	37.9	19.4	39.1	47.5	30.7	48.9	51.8	32.4	56.3	67.4
Detectron	35.7	54.7	38.5	19.5	39.9	47.5	30.7	49.1	52.0	32.0	56.9	68.0