An unofficial implementation of ICCV 2017 RetinaNet (Focal Loss).
Adapted from SSD implemented by zhreshold, the results still need to be tuned. Currently we use the PASCAL VOC mAP metric which measures under IoU threshold 0.5, not the COCO AP metric.
For PASCAL VOC and more details, one can generally refer to SSD implemented by zhreshold.
Tested on Ubuntu 16.04, python3.5, mxnet 1.1.0
Numpy, cv2 and matplotlib are required.
Backbone | Training data | Val data | Strategy | mAP | Note |
---|---|---|---|---|---|
ResNet-50 512x512 | VOC07+12 trainval | VOC07 test | OHEM | 76.0 | sgd, lr0.01 |
ResNet-50 512x512 | VOC07+12 trainval | VOC07 test | FL | 75.4 | sgd, lr0.01 |
ResNet-50 512x512 | COCO2017 train | COCO2017 val | OHEM | 40.2 | sgd, lr0.01 |
ResNet-50 512x512 | COCO2017 train | COCO2017 val | FL | 40.9 | sgd, lr0.01 |
Backbone | Training data | Val data | mAP | Note |
---|---|---|---|---|
ResNet-50 600 | VOC07+12 trainval | VOC07 test | 74.8 | sgd, lr0.001 |
ResNet-50 600 | COCO2017 train | COCO2017 val | 37.9 | sgd, lr0.003 |