Created by: guoshengCS

Add AnchorLayer for Faster R-CNN

AnchorLayer:

• Input: the last feature map(size), image(size), ground-truth bounding-boxes.
• Output: rpn_bbox_labels, rpn_bbox_targets.

With AnchorLayer's output, RPNLossLayer can be build just like MultiBoxLossLayer.

The AnchorLayer just rewrite the python code AnchorTargetLayer with Paddle framework. Would it be better if the AnchorLayer only generates the anchor location like PriorBoxLayer, and the labels and targets will be calculated in loss layer just like what the MultiBoxLossLayer does.