IT Trends/Conference, Faire (Experience)

[์„ธ๋ฏธ๋‚˜] Unsupervised Semantic Segmentation (22.03.18. Open DMQA Seminar)

DrawingProcess 2024. 7. 19. 10:30
๋ฐ˜์‘ํ˜•
๐Ÿ’ก ๋ณธ ๋ฌธ์„œ๋Š” 'Unsupervised Semantic Segmentation (22.03.18. Open DMQA Seminar)' ์„ธ๋ฏธ๋‚˜์„ ๋ฐ”ํƒ•์œผ๋กœ ์ž‘์„ฑํ•œ ๊ธ€์ด๋‹ˆ ์ฐธ๊ณ ํ•˜์‹œ๊ธฐ ๋ฐ”๋ž๋‹ˆ๋‹ค.

1. Introduction

2. Unsupervised Sementic Segmentation

Unsupervise Learning

  • ๋น„์ง€๋„ ํ•™์Šต์€ ์ •๋‹ต์ด ์ฃผ์–ด์ ธ ์žˆ์ง€ ์•Š์€ ์ƒํ™ฉ์—์„œ ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋งŒ์œผ๋กœ ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•๋ก 
  • ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋งŒ์œผ๋กœ ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•๋ก ์ด ์—ฐ๊ตฌ๋˜๊ณ  ์žˆ์Œ(Clustering, Autoencoder, GAN, Self-supervised, ...)

Unsupervise Learning for semantic segmentation

  • Sementic Segmentation์—์„œ ๋น„์ง€๋„ ํ•™์Šต์€ ์ •๋‹ต ๋งˆ์Šคํฌ๊ฐ€ ์ฃผ์–ด์ง€์ง€ ์•Š์€ ์ƒํ™ฉ์—์„œ ์ž…๋ ฅ ์ด๋ฏธ์ง€๋งŒ์œผ๋กœ ๋ชจ๋ธ์„ ํ•™์Šต
  • ๋‹จ, ํ•™์Šต์‹œ ํด๋ž˜์Šค์˜ ๊ฐœ์ˆ˜๋Š” ์ง€์ •๋˜์–ด ์ฃผ์–ด์ง€๋ฉฐ, ํ…Œ์ŠคํŠธ ๋‹จ๊ณ„์—์„œ๋Š” ์˜ˆ์ธก ๊ฒฐ๊ณผ์™€ ๋ ˆ์ด๋ธ”์„ ๋งคํ•‘ํ•˜์—ฌ ๋ชจ๋ธ์„ ํ‰๊ฐ€
    • ๋งคํ•‘ํ•˜๋Š” ๋ฐฉ๋ฒ•: Hungarian method(Kuhn, 1955) - Linear Assignment
  • ๋น„์ง€๋„ ํ•™์Šต ๊ธฐ๋ฐ˜์˜ Semantic Segmentation์€ IIC(Ji et al.) ์—ฐ๊ตฌ ์ดํ›„ ์ด๋ฅผ baseline์œผ๋กœ ํ•œ ์—ฐ๊ตฌ๋“ค์ด ์ง€์†๋˜๊ณ  ์žˆ์Œ
  • GAN ๊ธฐ๋ฐ˜์˜ ์—ฐ๊ตฌ๋„ ์กด์žฌํ•˜์ง€๋งŒ GAN ๊ธฐ๋ฐ˜์€ foreground์™€ background ๋งŒ์„ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์žˆ๋Š” ํ•œ๊ณ„๋ฅผ ์ง€๋‹˜: Labels4Free
  • [Leaderboard] Unsupervised Semantic Segmentation on COCO-Stuff: https://paperswithcode.com/sota/unsupervised-semantic-segmentation-on-coco-1

Unsupervise Learning for semantic segmentation: mutual information

  • IIC ์ดํ›„์˜ ์—ฐ๊ตฌ๋“ค์€ ๋Œ€๋ถ€๋ถ„ ๋ชจ๋ธ ํ•™์Šต์˜ objective๋กœ mutual information maximization์„ ์‚ฌ์šฉํ•จ
  • ๋”ฐ๋ผ์„œ Unsupervised semantic segmentation์˜ ์ตœ์‹  ์—ฐ๊ตฌ๋ฅผ ํŒŒ์•…ํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” mutual information์— ๋Œ€ํ•œ ์ดํ•ด๊ฐ€ ํ•„์š”ํ•จ.
  • ์ž…๋ ฅ ์ด๋ฏธ์ง€์™€ ์ฆ๊ฐ•๋œ ์ด๋ฏธ์ง€๋กœ positive pair๋ฅผ ๊ตฌ์„ฑํ•˜๊ณ  ์ด๋ฅผ ๋ชจ๋ธ์— ํ†ต๊ณผ์‹œํ‚จ ํ›„ ์ด์— ๋Œ€ํ•œ mutural information์„ ์ตœ๋Œ€ํ™” ํ•˜๋Š” ๋ชฉ์ ์‹์„ ํ†ตํ•ด ๋ชจ๋ธ์„ ํ•™์Šต

mutual information

  • ๋‘ ํ™•๋ฅ ๋ณ€์ˆ˜ ์‚ฌ์ด์˜ ์ƒํ˜ธ์˜์กด์„ฑ์„ ์ธก์ •ํ•œ ๊ฒƒ์œผ๋กœ, ํ•œ ๋ณ€์ˆ˜๋ฅผ ํ†ตํ•ด ์–ป์–ด์ง€๋Š” ๋‹ค๋ฅธ ํ•œ ๋ณ€์ˆ˜์— ๋Œ€ํ•œ ์ •๋ณด๋Ÿ‰์„ ์˜๋ฏธ
  • ๋‘ ํ™•๋ฅ ๋ณ€์ˆ˜์˜ ๋…๋ฆฝ์—ฌ๋ถ€๋ฅผ ์ธก์ •ํ•  ์ˆ˜ ์žˆ์–ด ์ƒ๊ด€๊ณ„์ˆ˜์™€ ์œ ์‚ฌํ•˜๊ฒŒ ๋‘ ๋ณ€์ˆ˜์‚ฌ์ด์˜ ๊ด€๊ณ„๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์ฒ™๋„๋กœ ์‚ฌ์šฉ

3. Mutual information maximization

Why Mutual information?

4. Method

1) IIC (Invarient Information Clustering)

1.1 

1.1.1 

 

์ฐธ๊ณ 

  •  
๋ฐ˜์‘ํ˜•