yolov3

Backbone network

v1 Structure

v2 Structure

v3 Structure , Will be output 3 The scale of the feature map , Namely 13x13 individual grid cell,26x26 individual grid cell and 52x52 individual grid cell,255 yes 3x85 Got , Every grid cell Will generate 3 individual anchor,85 yes 80+5,80 It's the number of categories ,5 Is the positional parameter plus confidence .13x13 Corresponding to the big receptive field , Predict large objects ;52x52 Corresponding to small receptive field , Predict small objects .

The green part is up sampling .v3 It realizes the integration of various scales .

Every grid cell Corresponding 3 individual anchor, Every anchor Corresponding to a prediction box , A prediction box corresponds to 85 Parameters ,


13x13 Responsible for predicting big goals , The three blue boxes correspond 3 individual anchor. For example, the Yellow frame of this dog , See where his center point falls grid cell Inside , Then look at this grid cell Which one? anchor With this yellow framed iou Maximum , Just use which one anchor.

And dimension box iou maximal anchor It is called positive sample ,iou It's big and nothing ,iou Less than 0.5 Is called negative sample .

The first is for positive samples , Fit the coordinates , The second is confidence and category , The third item is the confidence of negative samples .

During training , The parameters of the three scales are 10647 individual , Train them with the parameters of the callout box ,