resource:github上备份的包括ipad标注的pdf版本。
作者是马里兰大学的Venkataraman Santhanam, Vlad I. Morariu, Larry S. Davis等人,这写作水平给我看吐了,是不是不会好好说话!喜欢摆弄"高级"词汇是吧。
Summary:一篇比较菜的文章,贡献仅仅是模型结构设计,用一种循环分支+各分支上采样复用以提高scale的方法去做多任务(去噪、色彩恢复、人脸灯光增强)。创新点非常有限,还喜欢整点花里胡哨的词和晦涩难懂的句子,实在让人没有好感。
Key words: 模型结构设计、循环结构、参数复用
Rating: 2.0/5.0
Comprehension: 3.8/5.0
一张图总结模型结构:
要点:
the Recursively Branched Deconvolutional Network (RBDN) develops a cheap multi-context image representation very early on using an efficient recursive branching scheme with extensive parameter sharing and learnable upsampling. This multi-context representation is subjected to a highly non-linear locality preserving transformation by the remainder of our network comprising of a series of convolutions/deconvolutions without any spatial downsampling
讲一般的分类网络VGG和ResNet在图像回归任务上为什么不work的故事:
trade-off was skewed in favor of incorporating more context and subsequently reconstructing local correspondences from global activations
)同意。
activations from very early layers (which contain the bulk of the local correspondences) have a poor capability to model non-linearity, which limits the overall capacity of the network for modeling localized non-linear transformations
) -> 所以在模型前期就把不同尺度的特征提取出来正常的CNN结构,原文有些参数设置,每个Conv/DeConv后接ReLU和BN
主支感受野比较小,但是又不能用downsample,所以这里用了循环分支:
而所谓的可学习的上采样其实就是DeConv。
循环分支的构建有点抽象,不如直接去看3分支case:
循环构建的好处有两方面:
实验设置完全无聊,除了: