EffificientDet: Scalable and Effificient Object Detection
动机：
CC
CC
CC
CC
CC
CC
解题思路：
CC
CC
CC
CC
CC
BiFPN
CC
CC
CC
注
Figure 2 illustrates the design of feature networks.
(a) The FPN framework incorporates an upper pathway that integrates features across scales from levels 3 through 7 (P3-P7).
(b) PANet enhances this by adding a lower layer pathway.
(c) NAS-FPN leverages neural architecture search to identify an irregular feature topology and consistently applies identical blocks.
(d) Our BiFPN offers improved accuracy and efficiency trade-offs compared to existing methods.
CC
CC
CC
CC
CC
CC
CC
CC
CC
- Weighted Feature Fusion
CC
CC
Unbounded fusion:
CC
Softmax-based fusion:
CC
CC
Fast normalized fusion:
CC
Our finalized BiFPN combines two-way cross-scale connections with fast normalized fusion. As a specific instance, we elaborate on the two integrated features at level six for BiFPN as depicted in Figure 2(d):
CC
CC
EffificientDet
CC
CC
Figure 3 illustrates the EfficientDet architecture, which integrates a backbone network based on EfficientNet [36], coupled with a feature pyramid network (BiFPN), alongside a shared class and box prediction network. The BiFPN layers and class/box prediction layers are meticulously designed to be repeatedly employed under diverse resource constraints, as detailed in Table 1.
CC
CC
CC
CC
CC
- Backbone network
CC
CC
CC
- BiFPN network
Mathematically, the number of channels and layers in a BiFPN are determined by the following equation. Additionally, the network's depth is represented by Dbifpn (#layers), while its width is denoted by Wbifpn (number of channels).
【CC】也就是个检验公式，没啥好说的，下面作者自己也写了，也是试验出来的
- Box/class prediction network
We set their width equal to that of BiFPN (i.e., Wpred = Wbifpn), while gradually increasing the depth (#layers) through an equation.
- Input image resolution
In the BiFPN architecture, the input resolution must be divisible by 27, which equals 128. Thus, we linearly increase the resolutions through the use of an equation.
CC
CC
注
CC
CC
CC
CC
CC