Hi all
I found there are some small different setting masks
from yolov3.cfg
start at 6,7,8
[yolo]
mask = 6,7,8
from csresnext50-panet-spp.cfg
start at 0,1,2
[yolo]
mask = 0,1,2
is that mean csresnext50 should start from small anchors or just some mistake???
It just because yolov3 use FPN-like model (top-down), and csresnext50-panet-spp use PAN-like model (top-down -> bottom-up).
yolov3:
(start backbone) -> 1x -> 2x -> 4x -> 8x -> 16x -> 32x
-> (start top-down) -> 32x -> (predict 6,7,8) -> 16x -> (predict 3,4,5) -> 8x -> (predict 0,1,2)
csresnext50-panet-spp:
(start backbone) -> 1x -> 2x -> 4x -> 8x -> 16x -> 32x
-> (start top-down) -> 32x -> 16x -> 8x
-> (start bottom-up) -> 8x -> (predict 0,1,2) -> 16x -> (predict 3,4,5) -> 32x -> (predict 6,7,8)
Got it!!
Many thanks
Most helpful comment
It just because yolov3 use FPN-like model (top-down), and csresnext50-panet-spp use PAN-like model (top-down -> bottom-up).
yolov3:
(start backbone) -> 1x -> 2x -> 4x -> 8x -> 16x -> 32x
-> (start top-down) -> 32x -> (predict 6,7,8) -> 16x -> (predict 3,4,5) -> 8x -> (predict 0,1,2)
csresnext50-panet-spp:
(start backbone) -> 1x -> 2x -> 4x -> 8x -> 16x -> 32x
-> (start top-down) -> 32x -> 16x -> 8x
-> (start bottom-up) -> 8x -> (predict 0,1,2) -> 16x -> (predict 3,4,5) -> 32x -> (predict 6,7,8)