RCNet

Motivation:

Objectives:

Related Works:

Contribution:

Reverse Feature Pyramid:

Pyramid Pathway

1. Traditional FPN:

   • Top-Down pathway.

   • Pi = fi(Ci, Pi+1)

2. RevFP: 

   • Bottom-Up pathway + Local Top-Down pathway.

   • Adding a local top-down pathway between the 2 nearest feature maps.

   • Pi = fi(Ci, Ci+1, Pi-1)

Feature-guided Upsampling

1. Using spatial information from low-level feature maps and semantic information from high-level feature maps.

2. Weight Attention: 

   • Wi = Softmax(Conv3x3(Concat(Ci,Upsample(Ci+1))))

   • W'i = Wi * T/sqrt(di), where: di = N.o channels, T = constant.

   • C'i+1 = Upsample(Ci+1) * W'i

Dynamic Weighted Fusion

3. Weighted pre-fusion:

   • W'i = Sigmoid(Con1x1(Concat(Gap(Ci), Gap(C'i+1))))

   • W'i ∈ R1x1x1

   • P'i = Conv3x3(W'i * Ci + (1 - W'i) * C'i+1)

4. Weighted post-fusion

   • Pi = Conv3x3(W'i * P'i + (1 - W'i) * Pi-1)

Cross-Scale Shift Network:

• 2 feature maps at 2 close stages have a similar representative.

• Cross-scale Shift Network (CSN), have 2 modules: 

  1. Multi-scale shift.

  2. Dual global context.

Multi-scale Shift Module

Dual Global Context

• Scale context branch:

  1. spatial pooling

  2. channel context fusion

  3. scale pooling 

• Spatial context branch:

  1. scale pooling

  2. channel context fusion

  3. spatial pooling

Experimental Results:

Ablations:

References:

• https://viblo.asia/p/paper-explain-rcnet-kien-truc-fpn-dinh-cao-cho-object-detection-GAWVp0ZPV05 

1. • n2  n0

   • θ