Paper reading

[NarcissusInMirror]

阅读论文笔记

[NarcissusInMirror]

Network In Network

Lin et.al.

• 论文总的来说提出了这样一种网络结构：将卷积层后的激活函数改为一个MLP(Multi-layer-peceptron)，然后将top-layer改为GAP
• 卷积层的输出是要过激活函数的…… 一般是ReLu
• Global Average Pooling的思想出自于这篇文章，相当于限制输出层前的feature map数为类别数N，然后对每个feature map进行Global Average Pooling操作，即将每一张feature map上的值求平均，再通过softmax层即可。而Global Max Pooling就是取每一张feature map中的最大值。

[NarcissusInMirror]

Deep Residual Learning for Image Recognition

He et.al.

• ResNet论文。论文的核心思想是增加shortcut connection

• The degradation problem suggests that the solvers might have difficulties in approximating identity mappings by multiple nonlinear layers. 衰退问题意味着，程序很难不擅长通过多个非线性层构成一个恒等映射

• 以上这句话对于理解十分关键，作者最初的疑惑点就在于，网络不断加深时，即便是最差的情况下，后加的层什么都没学到，只是恒等映射输出，效果也不会比之前差，但是却出现了counterintuitive的现象，即层数加深后效果变差了。他得出的结论就是上面引用的一句话。

[NarcissusInMirror]

CBAM: Convolutional Block Attention Module

• 本篇论文主要的工作是在CNN中加入attention机制
• 设原始的特征形状为C×H×W
• channel attention module
  • channel attention是给每个通道不同的权值。当前的features经过channel attention的模块会输出一个C×1×1的向量，然后再与原始特征的每一层相乘image
• spatial attention module
  • spatial attention是给

[NarcissusInMirror]

Is object localization for free? –Weakly-supervised learning with convolutional neural networks

cited 456 times up to 19.6.20

• The fundamental challenge in visual recognition is modeling the intra-class appearance and shape variation of objects.

• • In this work, we focus on weakly supervised learning where only image-level labels indicating the presence or absence of objects are required.

  • 本篇文章中提到的弱监督学习就是只有一个image-level的标签，而不用bounding box给出具体的目标在哪里，其实毕设做的多标签分类，和很多图片分类问题都属于这种。文章中提到主要的困难是目标形态多样，被遮挡等等。其实之前做的都涉及到这些问题了，但没有考虑

• • The proposed method can also be seen as a variant of Multiple Instance Learning [21, 29, 54] if we refer to each image as a “bag” and treat each image window as a “sample”

  • 是多示例学习的变体

• • We build on the fully supervised network architecture of [37] that consists of five convolutional and four fully connected layers and assumes as input a fixed-size image patch containing a single relatively tightly cropped object.

  • 我认知中所默认的图片分类训练集其实包含这样的条件：每个物体基本填充满整张图片。之前并没有意识到这一点。

[NarcissusInMirror]

EXTD: Extremely Tiny Face Detector via Iterative Filter Reuse

cited 1 time up to 19.8.14

[NarcissusInMirror]

LFFD: A Light and Fast Face Detector for Edge Devices

cited 0 time up to 19.8.14

• 本论文使用的方法不使用anchor，是“一刀流”的

  The proposed method is anchorfree and belongs to the one-stage category.

• Face detection is the prerequisite to some face-related applications, such as face alignment and face recognition. 人脸检测是很多问题的前置步骤，先检测，再……

疑难

• 摘要部分

  Specifically, we rethink the importance of receptive field (RF) and effective receptive field (ERF) in the background of face detection. Essentially, the RFs of neurons in a certain layer are distributed regularly in the input image and theses RFs are natural “anchors”. Combining RF “anchors” and appropriate RF strides, the proposed method can detect a large range of continuous face scales with 100% coverage in theory.