经典论文

MODEL ARCHITECTURE

Alexnet

ImageNet Classification with Deep Convolutional Neural Networks.pdf

深度学习兴起的引子

VGG

Very Deep Convolutional Networks for Large-Scale Image Recognition

使用 3*3 卷积减少参数量的深层网络

GoogleNet，Inception系列

• Inception-V1: Going Deeper with Convolutions
• Inception-V2: Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift
• Inception-V3: Rethinking the Inception Architecture for Computer Vision
• Inception-V4: Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning, 即 inception-resnet
• Xception: Xception: Deep Learning with Depthwise Separable Convolutions

Densenet

Densely Connected Convolutional Networks

參差系列

何恺明个人主页http://kaiminghe.com/

• ResNet: Deep Residual Learning for Image Recognition 2015

  论文翻译

• ResNeXt: Aggregated Residual Transformations for Deep Neural Networks 2017

  pytorch

• SE-ResNet & SE-ResNeXt: Squeeze-And-Excitation Networks 2018

轻量级网络

• MobileNet-V1: MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
  • MobileNet: 可以牺牲少量性能来大幅降低网络的计算量
  • 将标准卷积转化为深度可分离卷积(depthwise conv + pointwise conv)，depthwise conv只在单一通道上进行卷积，提取特征，pointwise conv 为11N的卷积，将不同通道的特征连接起来，产生新的特征; 定义了两个超参 \alpha 和 \rho 用于进一步缩减网络，分别是宽度乘数和分辨率乘数，用于减少通道数和输入图像的大小，范围为(0,1);
  • MobileNet 可以用于分类、检测、识别等各个领域，适用于移动端和嵌入式端
• MobileNet-V2: MobileNetV2: Inverted Residuals and Linear Bottlenecks
• ShuffleNet-V1: ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices
• ShuffleNet-V2: ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design

LOSS FUNCTION

• AM-Softmax

  • Additive Margin Softmax for Face Verification
  • github caffe tensorflow

• A-Softmax

  • SphereFace: Deep Hypersphere Embedding for Face Recognition
  • github caffe

• L-Softmax

  • Large-Margin Softmax Loss for Convolutional Neural Networks
  • github caffe

• ArcFace

  • ArcFace: Additive Angular Margin Loss for Deep Face Recognition
  • github pytorch

• Triplet loss

  • In Defense of the Triplet Loss for Person Re-Identification
  • github

• Lovász-Softmax loss

  • The Lovász-Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks
  • github

Object Detection

• RCNN Rich feature hierarchies for accurate object detection and semantic segmentation
• Fast-RCNN Fast R-CNN
• Faster R-CNN Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
• RetinaNet-Focal Loss Focal Loss for Dense Object Detection
• Mask-RCNN Mask R-CNN
• YOLO
  • YOLOv3: An Incremental Improvement: YOLO-V3, 检测速度极大提升
• SSD: Single Shot MultiBox Detector:
  • SSD，目标检测
  • github pytorch caffe
• FaceBoxes: A CPU Real-time Face Detector with High Accuracy
  • 人脸检测的小网络，可以做到CPU实时，用到了SSD的思想
  • github
• Finding Tiny Faces: 小目边检测最好的方法，但是速度极慢
• An Analysis of Scale Invariance in Object Detection – SNIP: 小目标检测相关
• Generalized Intersection over Union: A Metric and A Loss for Bounding BoxRegression: 提出一种新的IOU计算方法

Semantic Segmentation

• FCNN
  • Fully Convolutional Networks for Semantic Segmentation： 用于语义分割的全卷积神经网络，开启使用卷积神经网络进行语义分割的先河;

Face Recognition

• 2019年最新人脸识别综述 Deep Face Recognition: A Survey
• 人脸识别的 one-shot learning 问题解决方案，earhian 推荐 One-shot Face Recognition by Promoting Underrepresented Classes

OTHERS

A guide to convolution arithmetic for deep learning

上采样的优点

Escaping From Saddle Points –Online Stochastic Gradient for Tensor Decomposition

batch的优点

Data Distillation: Towards Omni-Supervised Learning

数据蒸馏，一种使用无标签数据训练的全方位学习方法，在Kaggle等大数据竞赛中非常有用

IQA

No-reference Image Quality Assessment 相关论文，包括人脸姿态估计

DeepLearning-500-questions

四川大学深度学习500问，包含了深度学习数学基础、经典框架、常见问题等 github