Faster R-CNN

In fact, this is a more simplified version of simple-faster-rcnn-pytorch.

Improvements:

• use pytorch functions instead of numpy functions to generate anchors and targets. Be especially careful with these two functions: meshgrid where because they perform in different way in pytorch and numpy.
• clearer logic.

TO DO

• complete roi pooling
• add Resnet as backbone
• complete train and evaluate pipeline

Anchors Mechanism

ratios=(0.5,1.0,2.0) scales=(4.,8.,16.) base_size=16 stride=32  

to show the generated anchor box more clearly, only one set of base anchors is drawed.

ratios=(0.5,1.0,2.0) scales=(8.,16.,32.) base_size=16 stride=16  pos_iou_threshold=0.5

the green boxes are matched anchor boxes.

Model Architecture

for vgg16 as backbone, the architecture is shown in the figure below.