Lecture 11: Detection and Segmentation

CS231n

Lecture 11: Detection and Segmentation

Semantic Segmentation

Label each pixel in the image with a category label
Not differentiate instances, only care about pixels
最简单的想法：Sliding window，但是very inefficient
更好的方法：Fully Convolutional，但是也有问题：convolutions at original image resolution will be very expensive
解决方案：downsampling and upsampling inside the network
Upsampling: Transpose Convolution，之所以叫这个名字是因为从卷积的矩阵表示来看，这种操作就是把原kernel矩阵进行转置

Classification + Localization

对图像中的单个物体（默认图像中只有一个主要物体）进行分类和定位（图像中的物体不充满整张图像）

$$特征提取\Rightarrow\begin{cases}分类\\回归\mathrm{boundingbox}\end{cases}$$
然后loss相加

Human Pose Estimation

Object Detection

Sliding window

very computationally expensive

Region Proposals

Find “blobby” image regions that are likely to contain objects

R-CNN

Fast R-CNN

sharing features

Faster R-CNN

RPN

Detection without Proposals

YOLO / SSD

Dense Captioning

Object Detection + Captioning

Instance Segmentation

Mask R-CNN