A Novel Pyramid Network with Feature Fusion and Disentanglement for Object Detection

Figure 3

We propose two general structures of multiflow feature fusion methods: (d) 3-flow structure and (e) 5-flow structure. Native FPN (a), bidirectional FPN (b), and encoder-decoder FPN (c) are some other popular fusion methods. Red-dotted lines mean that they can be several operations including upsampling, downsampling, summing, and concatenation. The different directions of the red-dotted lines represent different information flows. Each solid black line presents an independent convolution. The red-dotted box represents the down-top subnetwork and the yellow-dotted box represents the top-down subnetwork.

(a)

(b)

(c)

(d)

(e)