Scale Adaptive Feature Pyramid Networks for 2D Object Detection

Figure 4

Fixed weight FPN using various ( : ) ratios (dashed lines) versus the SAFPN using square of L2-norm (solid lines) trained using the Pascal VOC 2007 + 2012 train dataset. Note that the original FPN [11] uses ( : ) = (0.5 : 0.5). The FRCNN-SAFPN adaptive weights.