Computer Vision

 

Full Stack Deep Learning 강의를 듣고 정리한 내용입니다.

📌AlexNet

• Deep layer with 8 layers
• First convnet winner
• Innovated with ReLU and Dropout
  • Dropout: some set of weights are set to zero randomly in training
• Heavy data augmentation(flip, scale, etc)
  • Fun fact: usually drawn as two parts, because the biggest GPU only had 3GB memory at the time
  • 그래서 model distributed training을 했어야 했다. (model이 두개의 다른 GPU에 위치한 상태)

📌ZFNet

• 거의 AlexNet과 동일 + just tweaking some parameters
• famous for deconvolutional visualizations
• each layers detect a certain type of image patch.
  • early layers: learn edge detection & texture detection(color detection)
  • later layers: detect parts of objects, e.g., ears, eyes, wheels, etc.

📌VGGNet

• More layers!
• Only use 3x3 convolutions & 2x2 max pools
• increased channel dimension with each layer
  - early layers: 64 channels
  • later layers: 512 channels
• has 138 million parameters
• most parameters are used in the later layers with the fully connected layers
  -사용되는 parameter를 제한하기 때문에 적은 수의 parameter로 계산할 수 있다-> 빠름!

📌GoogleNet

• =inception net
• VGG만큼 저렴하지만 3%정도의 parameter만 사용한다
• No fully-connected layers
• Is just a stack of inception modules( network module more creative than the standard)
• injected classifier outputs not only at the end but also in the middle
  • it let the network get gradient from the loss function at more spot than just at the end of the network

📌ResNet

• Very deep layer: 152 layers
• Down to 3.57% top-5 error(5%인 human performance보다 낮음!)
• most commonly used network now
• Problem: network가 깊어질수록 얕은 네트워크만큼의 성능을 내야하는데 그러지 않는 경우 발생
  - due to vanishing gradient
  -Solution: make an option to skip around layers
  • gradient가 vanish되는 경우 skip
• ResNet Variants
  - DenseNet: using more skip connections (skip connection을 건너뛰고 싶은 해당 layer만이 아니라 다른 모든 부분에 다 연결함)
  - ResNeXt: inception net + ResNet

📌SqueezeNet

• Focused on trying to reduce the number of parameters as much as possible
• Use constant 1x1 bottlenecking techniques
  • number of channels never expands

📌Overall Comparison

📌Localization, Detection, and Segmentation

• Classification: given an image, output the class of the object
• Localization: Do classification, but also highlight where the object is in the image
• Detection: given an image, output every object's class and location
• Segmentation: label every pixel in the image as belonging to an object or the background
  • Instance Segmentation: additionally differentiates between different objects of the same class
• Using networks for Localization
  
  - output bounding box coordinates(x1,y1,x2,y2) as well as the class of an object
  - class 결과를 제공할 때 사용하는 network와 동일한 네트워크를 사용해 마지막 단계에서 coordinate 예측값도 산출하게 한다.
• Using networks for Detection
  
  - 몇 개의 object가 있는지 모르는 상태라서 Localization에 사용한 방법을 사용할 수 없음
  - Solution: slide a classifier over the image(at multiple scales)
  - VERY computationally expensive, but 해결방법 있음!

📌Non-maximum Supression(NMS) and IOU

• Non-maximum Supression(NMS): When multiple bounding boxes overlap, you should keep the one with the highest score and remove all the others.
• Intersection over Union(IOU): most common metric for localization quality

📌YOLO/SSD

• YOLO(You Only Look Once)
  
  1. Put a fixed grid over an image, and within the grid find objects
  2. Output class and box coordinates
  3. Run non-maximum supression
  - nice & fast, and is in active development!
• Microsoft COCO: Common Objects in Context
  
  - 이 dataset을 가지고 YOLO의 성능을 평가함.
  - 330,000 images & 1.5 million object instances & 80 categories & some captions

📌Region Proposal Methods

• 이제까지는 이미지의 모든 부분을 관찰함.
  • 그러지 말고 중요해 보이는 부분만 관찰하면 어떨까? (look only at regions that seem interesting)
• R-CNN(Region-CNN)
  1. Using external(non-deeplearning) methods to find regions
  2. Use AlexNet on regions
  3. Predict both class and bounding box(coordinates)
• Faster R-CNN
  -Used convnet for the Regional Proposal Network
  • Region Proposal Network(RPN): a fully convolutional method for scoring a bunch of candidate windown for "objectness"
    

    

  • Faster (because everything is done in the convnet!)
  • Four losses total: classifier and bbox regression for both RPN and object classifier
• Mask R-CNN
  -Each region goes in not only the classification but also the segmentation step
  • Regions go through a couple of non-downsampling convolutions

📌Adversarial Attacks

• Convnets can be brittle in unexpected ways
• Convnet 공격 방법
  • Add noise to an image
    

    

  • Add real things in real world such that when you take an image of it, it messes up the network
    

    

    
    - 자율주행자동차 운행에 문제
• Who would win?
  - Detect 방법: try to find inputs that push the gradient of the network towards some class very strongly -Defend 방법:
  - adversarial example을 같이 traing하기(doesn't work well)
  - Smooth the class decision boundaries(= Defensive distillation)