Deep lab家族ppt.pdf一个官方ppt，涵盖了Deep LabV1V2V3的概括总结，能够

文件名称: Deep lab家族ppt.pdf

所属分类: 深度学习

开发工具:

文件大小: 2mb

下载次数: 0

上传时间: 2019-07-08

提供者: weixin_********

下载 (2mb)

不能下载？报告错误

详细说明：一个官方ppt，涵盖了Deep LabV1V2V3的概括总结，能够丰富语义分割论文内容Semantic segmentation Semantic segmentation Semantic segmentation is understanding an image at pixel level i. e, we want to assign each pixel in the image an object class Partitioning an image into regions of meaningful objects Assign an object category label grass building water sheep car COW oa d Semantic segmentation Why semantic segmentation? Autonomous driving Medical purposes We will focus on 3 papers C3 DeepLabV background DeeplabV2 Deeplabv3 horse semantic segmentation DeepLabvl& DeepLabv2 Use dcnn for classification to generate a rough prediction of segmentation (smooth, blurry heat map) Refine prediction with conditional random field (CRF) 日令 Image cNn output CRF output DCNN Score maps per Class Dog DCNN Cat Background per pixel What happens in standard donne Striding- put size Pooling- to small translations of the input DeepLab solution Atrous convolution CRFS(Conditional Random Fields DCNN-AtroUs(Holes) Remove the last 2 pooling layers Up-sample the original filter by a factor of the strides rate = 2) Standard convolution responses at only 1/4 of the image positions ■ Convolve image with a filter with holes'> responses at all image positions corva stride 2 门[ 2 almug tuirvululislI stade=1 DCNN-AtroUs(Holes) Small field-of-view> accurate localization Large field-of-view> context assimilation Effective filter size increases Both the number of filter parameters and the number of operations per position stay constant DCNN-AtroUs(Holes) The authors found a good efficiency /accuracy trade-off, Using atroUs convolution to increase by a factor of 4 the density of computed feature maps, followed by bilinear interpolation(factor 8) Input Input Aeroplane DCNN Aeroplane coarse Score m Dee Coar se Score Imap Atrous Convolution Corivoluliarial Neural Network A△ △△△ Final output Fully Connectec CRF Bi-linear Inter polation Final output Fully Connected CRF Bi-linear Interpolation the proposed approach converts image classification networks into dense feature extractors without requiring learning any extra parameters AtroUs(Holes)-Some Results Kernel Rate Fov Params Speed bef/aft CRF 7×7 448 224 134.3M 1.44 64.38/67.64 4×4 128 65.1M 2.90 59.80/63.74 4×4 224 65.1M 2.90 63.41/67.14 3×3 12 224 20.5M 4.846225/67.64

(系统自动生成,下载前可以参看下载内容)