paper--Faster RCNN

Some understanding about the details in Faster RCNN, based on the codes in tensorflow.

Big Question

reason

background

Region Proposal Network

Some numbers

• The number of anchor boxes for one anchor target \(k = scale \times ratios\),
• The number of anchor boxes for one feature layer (which has \(W \times H\) grids), will get $ W H k $ anchor boxes. Every grid in the feature map (the output of a popular CNN without FC layers) will have \(k\) anchor boxes.
• Not like the ROI method, the size of features are fixed, but anchor boxes are rescaled by \(k\) regressors.

Experiments

prove

• The top-ranked RPN proposals are accurate.
• NMS does not harm the detection mAP and may reduce false alarms.

Construct

Add loss

Problems using it processing typhoon data

• Does NMS lead to loss of typhoon? Not really, the texture of typhoon is obvious in image.