Computer Vision

Computer Vision (CV) Tasks

• Classification
• Classification + Localization
• Object Detection
• Instance Segmentation

Object Localization: Coordinate prediction

Sliding Window

Object Localization

Classification & Localization

Detection

Sliding Window + Classification:

Regioning

• Sliding Window Problem: Need to test many positions and scales, and use a computationally demanding classifier

• Solution: Only look at a tiny subset of possible positions

  • Regioning => propose image regions that are likely to contain objects
  • Classify individual regions and correct regions
  • R-CNN -> Fast R-CNN -> Faster R-CNN

R-CNN

• Propose approx. 2k different regions (bounding boxes) for image classification
• For each box, do image classification with CNN
  • Discard unlikely boxes
• Refine bounding boxes with regression

Fast R-CNN

• 9x faster training, 213x faster test time
• R-CNN is not end to end (first train softmax classifier, use that for training bounding box regressor)
• Similar to R-CNN
  • Apply Region Proposals on feature map result of applied CNN to input image
  • Reshape region proposals on feature map into fixed size
  • Feed into FC layer

Faster R-CNN

• Both R-CNN and R-CNN rely on Selective Search for region proposals -> most time consuming part 🤪
• Use a seperate Network for predicting the regions of interest 💪

YOLO

• You Only Look Once: Unified Real-Time Object Detection

• „Simple network“, directly from pixels to bounding box / object detection / class prediction

Image Segmentation

• Grouping Pixels into regions that belong to same properties
• Eg: Segmenting an Image into meaningful objects

Semantic Segmentation

Sliding Window

• Label each pixel in image with a category label

• Don‘t differentiate instances, only care about pixels

• => just extract small patches from an image and classify center pixel with a normal CNN classifier

  

• Problem: very inefficient

Fully convolutional

• Keep the network as an end to end convolutional Neural Network

• Predictions are made for all pixels at once

  

• Convolutions at original image resolution are very expensive

Reference

• Object Detection for Dummies Part 3: R-CNN Family