1) How (Is it possible) to combine Fast-RCNN (2-stage) and YOLO (1-stage)?
2) Why with the addition of anchor boxes we changed the resolution to 416 x 416? Why using anchor boxes we get a small decrease in accuracy? How does using anchor boxes decouple the class prediction mechanism from the spatial location? How do you guys feel about CenterNet (also a single-stage detector, but without anchor box)?
3) Why if we use standard k-means with Euclidean distance larger boxes generate more error than smaller boxes? How to derive d(box, centroid) = 1 - IOU(box, centroid)?
4) How to derive Pr(object) * IOU(b, object) = σ(to) in Yolo v2 ? Why is this expression not used in Yolo v3?
5) Why apply a Sigmoid function to tx and ty? Why apply an exponential function to pw and ph?
6) From yolov3-spp.cfg, I did not see anything about 3 different scales. Could anyone advise?
7) Why does YOLO v3 tensor size need to be multiplied by N*N? What is represented by N?