โ›บ DAY 32. ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜ ์‹ฌํ™”์™€ Semantic Segmentation

thumbnail

๐Ÿ™Œ์€ QnA์— ์žˆ๋Š” ์งˆ๋ฌธ-๋‹ต๋ณ€์„ ํ†ตํ•ด ์–ป์€ ์ง€์‹์„ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค.

๐Ÿ‘‰ ํ”ผ์–ด ์„ธ์…˜

๊ธฐ๋ก

  • ์˜ค๋Š˜์˜ Hot Issue๐Ÿ”ฅ๋Š” Semantic Segmentation์ด์—ˆ๋‹ค. ํŠนํžˆ Deconvolution ๋ถ€๋ถ„๊ณผ DeepLab ๋ชจ๋ธ ๊ตฌ์กฐ๊ฐ€ ์ดํ•ด๊ฐ€ ์•ˆ ๋˜์„œ ์ด ๋ถ€๋ถ„๊ณผ ๊ด€๋ จ๋œ ์งˆ๋ฌธ์ด ๋งŽ์ด ๋‚˜์™”๋‹ค. ํŠนํžˆ Deconvolution ๋ถ€๋ถ„์ด ์ง๊ด€์ ์œผ๋กœ ์ดํ•ด๊ฐ€ ๋˜์ง€ ์•Š์•„์„œ ๋งŽ์€ ์‹œ๊ฐ„์„ ํ• ์• ํ•ด ์ด์•ผ๊ธฐ๋ฅผ ํ•˜์˜€๋‹ค.
  • Deconvolution์˜ ๋ฌธ์ œ์ ์€ ํ•„ํ„ฐ์˜ ํฌ๊ธฐ๊ฐ€ Stride๋กœ ๋‚˜๋ˆ„์–ด ๋–จ์–ด์ง€์ง€ ์•Š์œผ๋ฉด ๊ฒฐ๊ณผ ์ด๋ฏธ์ง€๊ฐ€ ๊ผญ ์ฒด์Šค๋ณด๋“œ ์ฒ˜๋Ÿผ ๋‚˜์˜ค๋Š” ๊ฒƒ์ด๋‹ค. ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ํ•„ํ„ฐ์˜ ํฌ๊ธฐ์— Stride๊ฐ€ ๋‚˜๋ˆ„์–ด ๋–จ์–ด์ง€๋„๋ก Stride์˜ ํฌ๊ธฐ๋ฅผ ์ˆ˜์ •ํ•˜๊ฑฐ๋‚˜ ์ด๋ฏธ์ง€๋ฅผ Interpolation์œผ๋กœ Upscaling์‹œํ‚ค๊ณ  ๋ณ€ํ˜•๋œ ์ด๋ฏธ์ง€๋ฅผ Deconvolution์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•œ๋‹ค.
  • Dilated Convolution์€ Deconvolution์ด ์•„๋‹Œ Convolution์ด๋‹ค!(๊ทธ๋ฆผ์ด ๋น„์Šทํ•ด์„œ ํ—ท๊ฐˆ๋ฆฐ๋‹ค๐Ÿ˜ฅ) ๊ทธ๋ฆฌ๊ณ  Receptive Field๋Š” ๋‚ด์ ๋  ๋•Œ ์‚ฌ์šฉํ•˜๋Š” ๊ฐ’์˜ ๋ฒ”์œ„๊ฐ€ ์•„๋‹Œ ํ•„ํ„ฐ๊ฐ€ ๋ฎ๋Š” ์ž…๋ ฅ์˜ ํฌ๊ธฐ์ด๋‹ค. ํ—ท๊ฐˆ๋ฆฌ์ง€ ๋ง ๊ฒƒ!

Table of Contents

โœ DAY 13. Convolutional Neural Network์— ์žˆ๋Š” ๋‚ด์šฉ ์™ธ์˜ ๊ฒƒ๋งŒ ์ •๋ฆฌํ•ฉ๋‹ˆ๋‹ค!

์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜ ์—ญ์‚ฌ: GoogLeNet, ResNet, Beyond ResNet

๐Ÿ‘€ CNN ํ†บ์•„๋ณด๊ธฐ ๋‹ค์‹œ ๋ณด๊ธฐ

GoogLeNet

32 googlenet

ResNet

Beyond ResNet

CNN Backbone์œผ๋กœ ์ ์ ˆํ•œ ๋ชจ๋ธ์€?

๐Ÿ‘€ ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜ ์—ญ์‚ฌ๋ฅผ ํ•œ ๋ˆˆ์— ๋ณด๊ธฐ

32 image classification history

Semantic Segmentation

๐Ÿ‘€ Semantic Segmentation ๋‹ค์‹œ ๋ณด๊ธฐ

FC Layer๋ฅผ 1ร—1 Conv Layer๋กœ!

Upsampling

Transposed Convolution์˜ ๋ฌธ์ œ์ 

ํ•ด๊ฒฐ์ฑ…: Interpolation + Convolution

Skip-connection

Hypercolumns

FCN ๋ชจ๋ธ๋“ค: U-Net, DeepLab

U-Net

DeepLab

References


๐Ÿ‘‹@์ฝ”๋”ฉํ•˜๋Š”ํŽญ๊ท„
ํŒŒ์ด์ฌ๊ณผ ์›น์— ๊ด€์‹ฌ ๋งŽ์€ ์ปด๊ณต ์ „๊ณต์ž

GitHubInstagram