计算机视觉的主要任务和进展

计算机视觉任务、可视化和理解

Semantic Segmentation

训练集: 给每个像素点 打上语义类别标签

测试 : 给图片的每个像素点分类

思路一: Sliding Window ,上下文提取特征

Learning Hierarchical Features for Scene Labeling(TPAMI 2013)

Recurrent Convolutional Neural Networks for Scene Labeling(ICML 2014)

问题:效率很低,没有利用重叠的patchs的共享features.

思路二:全卷积

idea1: 提取上下文空间特征,不下采样从而保证输出和输入的shape一致$\Rightarrow$ 太贵了

Idea2:先下采样然后再上采样来keep shape

• Fully Convolutional Networks for Semantic Segmentation(CVPR 2015)

• Learning Deconvolution Network for Semantic Segmentation(ICCV 2015)

下采样 : 先Pooling 再 Stride Convolution

上采样 : Unpooling

可学习的版本: 反卷积(Transpose Convolution)

思路三: U-Net

$\color{red} 别关心实例,只关注像素 $​

• 先downsampling 再upsampling

物体检测=Classification + Localization

但是对于多物体检测, Each image needs a different number of outputs!

思路一: 将图片分成多个patch然后分别进行单物体检测

Problem: Need to apply CNN to huge number

of locations, scales, and aspect ratios, very

computationally expensive!

思路二: Region Proposals: Selective Search

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

● Relatively fast to run; e.g. Selective Search gives 2000 region

proposals in a few seconds on CPU

思路三: R-CNN

Rich feature hierarchies for accurate object detection and semantic segmentation(CVPR 2014)

Modern Method

Single-Stage Object Detectors: YOLO / SSD / RetinaNet

Yolo(CVPR 2016)

DETR:Object Detection with Transformers(ECCV 2020)