上采样下采样 #8
Description
关于代码和论文对应的问题：
查阅您的论文和代码后我认为您提出的上采样方法SNI在代码中是
def upsample_add(self, x, y):
'''Upsample and add two feature maps.
Args:
x: (Variable) top feature map to be upsampled.
y: (Variable) lateral feature map.
Returns:
(Variable) added feature map.
Note in PyTorch, when input size is odd, the upsampled feature map
with F.upsample(..., scale_factor=2, mode='nearest')
maybe not equal to the lateral feature map size.
e.g.
original input size: [N,,15,15] ->
conv2d feature map size: [N,,8,8] ->
upsampled feature map size: [N,,16,16]
So we choose bilinear upsample which supports arbitrary output sizes.
'''
,,H,W = y.size()
# return F.upsample(x, size=(H,W), mode='bilinear') + y
return F.upsample(x, size=(H, W), mode='bilinear')*0.25 + y
但是关于下采样的方法ESD，我没有找到相关代码。而且根据您论文的描述，您是一个独立的特征金字塔，下采样过程会出现在哪个部分。根据您的论文描述，不同层级特征图通过骨干网络输出c2c3c4c5，后通过上采样并叠加得到p2p3p4p5。ESD作用在哪个位置呢。