RetinaNet-mxnet

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.

Demo Results

Differences from SSD

• We build FPN (P3 to P7) to replace the "multi_layer_feature" function;

• We build cls_subnet and bbox_subnet in the "multibox_layer" function, and the bias is initialized according to the Focal Loss paper (only for FL strategy

  );

• We use the anchor setting in the focal loss paper which is tested on COCO, but the best setting for PASCAL VOC still needs to be tuned;

• We adopt the focal loss operator by eldercrow;

• We support converting COCO2017 data to rec format for training and validation.

Usage

• Download COCO2017 data and annotations;

• Run tools/prepare_coco.sh to pack into rec format, after configuring your own paths;

• Run train-COCO2017.sh after configuring your own paths and hyperparamters.

For PASCAL VOC and more details, one can generally refer to SSD implemented by zhreshold.

Environment

Tested on Ubuntu 16.04, python3.5, mxnet 1.1.0

Numpy, cv2 and matplotlib are required.

mAP result

Baseline Faster RCNN