1. 程式人生 > >CS231n筆記 Lecture 11, Detection and Segmentation

CS231n筆記 Lecture 11, Detection and Segmentation

des nds input ply instance oca output height tran

Other Computer Vision Tasks

  • Semantic Segmentation. Pixel level, don‘t care about instances.
  • Classification + Localization. Single object.
  • Object Detection. Multiple object.
  • Instance Segmentation. Multiple object.

Semantic Segmentation

Simple idea: sliding window, crop across the whole image, and ask what the center pixel is. Expensive.

Fully Convoltional (Naive) : let the network to learning all the pixels at once, keep the spacial size, convolutions at original image resolution, expensive.

Fully convolutional: Design network as a bunch of convolutional layers, with downsampling and upsampling inside the network!

  • Downsampling: Pooling, strided convolution
  • Upsampling: Unpooling (nearest neighbor, bed of nails, max unpooling in symetrical NN), Transpose convolution (multiply the filter by the pixels on the input, use stride and pad to impose the value on the output).
    技術分享圖片

  技術分享圖片

Classification + Localization

Get class scores and box coordinates from the CNN, treat localization as a regression problem, we have 2 loss!

Aside: Human Pose Estimation, for different position, multitask loss.

Object Detection

Since we have different numbers of objects present, it‘s impossible to use regression. Naively, sliding window.

R-CNN: Based on tranditional techniques in CV, gives thousands proposal region, much better.

Fast R-CNN: Region crop after ConvNet.

Faster R-CNN: Proposal Region Network.

YOLO/SSD: base grids.

Instance Segmentation

Mask R-CNN

CS231n筆記 Lecture 11, Detection and Segmentation