Scale Adaptive Feature Pyramid Networks for 2D Object Detection

Figure 5

Examples of fixed weights using FPN (a) and adaptive weights using SAFPN (b). SAFPN gave large weight to high resolution feature map c₃ in (B) and (D), which helps detection of small objects. For image (F) containing a large object, SAFPN gave large weight to low resolution feature map c₅. (a) A “cow” far in the background is not detected . (b) A small “cow” far in the background is detected. Smaller scales (level 2 and 3) have higher weights for this image . (c) Small cow near the center of image is not detected . (d) Small “cow” near the center of the image is detected. Smaller scales (levels 2 and 3) have higher weights for this image . (e) This image contains easy-to-detect large object, an airplane . (f) With a large object, SAFPN weighs larger scale (i.e., lowest resolution) feature map at level 5 .

(a)

(b)

(c)

(d)

(e)

(f)