Notice
Recent Posts
Recent Comments
ยซ   2024/11   ยป
์ผ ์›” ํ™” ์ˆ˜ ๋ชฉ ๊ธˆ ํ† 
1 2
3 4 5 6 7 8 9
10 11 12 13 14 15 16
17 18 19 20 21 22 23
24 25 26 27 28 29 30
Tags
more
Archives
Today
Total
๊ด€๋ฆฌ ๋ฉ”๋‰ด

Hello Potato World

[ํฌํ…Œ์ดํ†  ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ] Learning Data Augmentation Strategies for Object Detection ๋ณธ๋ฌธ

Paper Review๐Ÿฅ”/Data Augmentation

[ํฌํ…Œ์ดํ†  ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ] Learning Data Augmentation Strategies for Object Detection

Heosuab 2020. 12. 9. 23:03

 

โ‹† ๏ฝก หš โ˜๏ธŽ หš ๏ฝก โ‹† ๏ฝก หš โ˜ฝ หš ๏ฝก โ‹† 

[Data Augmentation/Object Detection paper review]

 

Classification์—์„œ์˜ Data Augmentation์€ ๋งŽ์ด ๋‹ค๋ค„๋ดค์ง€๋งŒ Object Detection์—์„œ๋Š” ๊ตฌ์ฒด์ ์œผ๋กœ ์–ด๋–ป๊ฒŒ ์ด๋ค„์ง€๋Š”์ง€ ๋ฌธ๋“ ๊ถ๊ธˆํ•ด์ ธ์„œ ์„œ์นญํ•˜๋‹ค๊ฐ€ ์ฝ๊ฒŒ ๋œ ๋…ผ๋ฌธ. ์ƒ๊ฐ๋งŒํผ ์„ ํ–‰์—ฐ๊ตฌ๊ฐ€ ๋งŽ์ด ์ด๋ค„์ง€์ง„ ์•Š์€ ๊ฒƒ ๊ฐ™๊ณ  detection ์™ธ์—๋„ ๋งŽ์€ ๋‚ด์šฉ์„ ๊ณต๋ถ€ํ•ด์•ผ ์ •ํ™•ํ•˜๊ฒŒ ์ดํ•ดํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ ๊ฐ™๋‹ค.

 

 


Learning Data Augmentation Strategies for Object Detection


 

Data Augmentation(๋ฐ์ดํ„ฐ ์ฆ๊ฐ•)์€ ํ•™์Šต ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•œ ์ƒํ™ฉ์ด๋‚˜, ํ•™์Šต ๋ฐ์ดํ„ฐ๋ฅผ ๋Š˜๋ ค์„œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋†’์ด๊ณ  ์‹ถ์„ ๋•Œ ์‚ฌ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์ด๋‹ค. ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•œ ๊ฒฝ์šฐ ๋ฐ์ดํ„ฐ์…‹์˜ ํŠน์ง•๋“ค์„ ์ž˜ ์žก์•„๋‚ด์ง€ ๋ชปํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ Overfitting, Underfitting์— ๋น ์ง€๊ธฐ ์‰ฝ๋‹ค. ๊ทธ๋ฆผ์—์„œ ๋ณด๋Š”๊ฒƒ๊ณผ ๊ฐ™์ด ์›๋ณธ ์ด๋ฏธ์ง€์— ์ธ์œ„์ ์ธ noise๋‚˜ ๋ณ€ํ™”๋ฅผ ์ฃผ์–ด ๋ฐ์ดํ„ฐ์˜ ์–‘์„ ์ฆํญ์‹œํ‚ค๋Š”๋ฐ ํšŒ์ „, ์ƒ‰๋ณ€ํ™˜, ์ž˜๋ผ๋‚ด๊ธฐ, ์ผ๋ถ€ ํ”ฝ์…€ ๋ณ€ํ™˜, ์Šค์ผ€์ผ๋ง, ๋’ค์ง‘๊ธฐ, ๋ฐ๊ธฐ ๋ณ€ํ™” ๋“ฑ... ์—ฌ๋Ÿฌ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์ด ์žˆ์„ ์ˆ˜ ์žˆ๋‹ค.

[Figure 1: Data Augmentation in Classification]

Classification์—์„œ๋Š” ์ด๋ฏธ Data Augmentation์ด ํ•„์ˆ˜์ ์ด๊ณ , ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋†’์ด๋Š”๋ฐ ํฐ ๋„์›€์„ ์ค€๋‹ค๋Š” ๊ฒŒ ๋งŽ์ด ์ž…์ฆ์ด ๋˜์—ˆ์ง€๋งŒ Object Detection์—์„œ์˜ Data Augmentation์€ ์•„์ง ์˜จ์ „ํ•˜๊ฒŒ ์—ฐ๊ตฌ๋˜์ง€ ์•Š์•˜๋‹ค. ๋…ผ๋ฌธ ๋‚ด์—์„œ ์–ธ๊ธ‰๋œ ์›์ธ์€ ๋‘๊ฐ€์ง€ ์ •๋„๊ฐ€ ์žˆ๋Š”๋ฐ

  1. Object detection์—์„œ๋Š” image annotation, bounding box ๊ฐ’ ๋ณ€ํ™”๋ฅผ ํ•จ๊ป˜ ์ค˜์•ผํ•˜๋Š” ์ถ”๊ฐ€์ ์ธ ์—ฐ์‚ฐ์ด ํ•„์š”ํ•˜๋‹ค.
  2. Classification์— ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์–ดํ„ฐ์…‹๋ณด๋‹ค detection์„ ์œ„ํ•œ ๋ฐ์ดํ„ฐ์…‹์˜ example์ด ๋” ์ ๋‹ค.

[Figure 2: Data Augmentation in Object Detection]

Figure2์—์„œ ๋ณผ ์ˆ˜ ์žˆ๋“ฏ์ด, Object Detection์€ ๊ฐ์ฒด์˜ ์œ„์น˜๋ฅผ bounding box์— ๊ธฐ๋ฐ˜ํ•ด์„œ ๊ฒ€์ถœํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋ณ€ํ™˜ ๋ฐฉ๋ฒ•์— ๋”ฐ๋ผ์„œ bounding box์˜ ์ขŒํ‘œ๊ฐ’๋„ ํ•จ๊ป˜ ๋ณ€ํ™˜ํ•ด์•ผํ•  ์ˆ˜๋„ ์žˆ๊ณ , ์—ฌ๋Ÿฌ๊ฐ€์ง€ ๊ฒฝ์šฐ๋ฅผ ๊ณ ๋ คํ•ด์•ผ ํ•˜๊ธฐ ๋•Œ๋ฌธ์— Classification๋ณด๋‹ค ๋ณต์žกํ•œ ๊ณผ์ •์„ ๋”ฐ๋ฅธ๋‹ค.

 

๋…ผ๋ฌธ์—์„œ๋Š” Classification์—์„œ ์‚ฌ์šฉํ•˜๋˜ ์ด๋ฏธ์ง€ ๋ณ€ํ™˜ ๋ฐฉ์‹(operations)๋ฅผ ๊ทธ๋Œ€๋กœ ๋นŒ๋ ค์˜ค๋˜, detection์— ๋งž๊ฒŒ ๋ณ€ํ™˜ํ•˜๊ธฐ ์œ„ํ•ด 3๊ฐ€์ง€ ๊ฒฝ์šฐ์˜ operations๋ฅผ ์š”์•ฝํ–ˆ๋‹ค.

  1. Color operations : ์ด๋ฏธ์ง€์˜ color๊ฐ’์„ ๋ณ€ํ™˜ํ•˜๋˜, bounding box์˜ ์œ„์น˜ ์ขŒํ‘œ๊ฐ’์—๋Š” ๋ณ€ํ™”๋ฅผ ์ฃผ์ง€ ์•Š๋Š”๋‹ค.
    • Equalize, Contrast, Brightness.....๋“ฑ(์ฃผ๋กœ PIL๋ฅผ ์‚ฌ์šฉํ•œ ๋ณ€ํ™˜)
  2. Geometric operations : ์ด๋ฏธ์ง€์˜ ์œ„์น˜์ •๋ณด๋ฅผ ๋ฐ”๊พธ๋ฉฐ(๊ฐ์ฒด์˜ ์œ„์น˜์— ๋ณ€ํ™”๊ฐ€ ์ƒ๊น€), bounding box annotation์˜ ์œ„์น˜๋‚˜, ์‚ฌ์ด์ง€๋ฅผ ๊ฐ™์ด ๋ณ€ํ™”์‹œํ‚จ๋‹ค.
    • Rotate, ShearX, TranslationY.....๋“ฑ
  3. Bounding box operations : ์ด๋ฏธ์ง€ ๋‚ด์—์„œ bounding box annotation์ด ์žˆ๋Š” ๋ถ€๋ถ„์˜ ํ”ฝ์…€๋งŒ ๋ณ€ํ™”์‹œํ‚จ๋‹ค.
    • BBox_Only_Equalizae, BBox_Only_Rotate, BBox_Only_FlipLR.....๋“ฑ

์ด operations๋Š” training ๊ณผ์ •์—์„œ ์‚ฌ์šฉํ•˜๋ฉฐ(test์ค‘์—๋Š” ์‚ฌ์šฉํ•˜์ง€ ์•Š๊ณ  ๊ธฐ์กด๊ณผ ๋™์ผํ•˜๊ฒŒ ์œ ์ง€ํ•ด์•ผํ•œ๋‹ค.) ํ•˜๋‚˜์˜ ์ด๋ฏธ์ง€์— ๋Œ€ํ•ด ์ˆœ์ฐจ์ ์œผ๋กœ ์—ฌ๋Ÿฌ ๊ฐœ ์ ์šฉ๋˜๋Š”๋ฐ, ์–ด๋–ค ๋ณ€ํ™˜์„ ์–ด๋–ค ์ˆœ์„œ๋กœ ์‚ฌ์šฉํ• ์ง€ Search method๋ฅผ ํ†ตํ•ด ์ตœ์ ํ™”์‹œํ‚จ๋‹ค.

 

[Figure 3: Search Space]

๊ฐ๊ฐ N๊ฐœ์˜ ์ˆœ์ฐจ์ ์ธ transformation operands๋ฅผ ๊ฐ–๋Š” K๊ฐœ์˜ sub-policy๋ฅผ ํ•™์Šตํ•˜๊ณ  training ๊ณผ์ •์—์„œ ๊ฐ ์ด๋ฏธ์ง€์— ์ ์šฉ๋  policy๊ฐ€ ๋žœ๋ค์œผ๋กœ ์„ ํƒ๋œ๋‹ค. Figure3๋Š” K=5, N=2์ผ ๋•Œ์˜ Search Space์ด๊ณ , ๊ฐ๊ฐ์˜ operands๋Š” ๋‹ค์Œ ๋‚ด์šฉ์— ํ•ด๋‹นํ•˜๋Š” ์ด 3๊ฐœ์˜ ํŒŒ๋ผ๋ฏธํ„ฐ(predictions)๋ฅผ ๊ฐ€์ง„๋‹ค.

  1. ์–ด๋–ค image transformation๊ฐ€ ์„ ํƒ๋๋Š”์ง€
  2. transformation์ด ์ ์šฉ๋  ํ™•๋ฅ  M
  3. transformation์ด ์ ์šฉ๋  ํฌ๊ธฐ L (ex ๋ช‡๋„ ํšŒ์ „?)

์ด ์กฐ๊ฑด ๋‚ด์—์„œ ์ข‹์€ sub-policy๋ฅผ ๊ณ ๋ฅด๋Š” ํ™•๋ฅ ์€ ์‚ฌ์ „ ์—ฐ๊ตฌ์— ์˜ํ•ด(๋‹ค๋ฅธ ๋…ผ๋ฌธ์ด ๋ถ€๋ก์œผ๋กœ ์‹ฌ์–ด์ ธ์žˆ๋Š”๋ฐ ์ถ”ํ›„์— ๋ด์•ผํ•  ๊ฒƒ ๊ฐ™๋‹ค) ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์‹์œผ๋กœ ์ •์˜๋˜๊ณ , sub-policy๊ฐ€ 5์ผ ๊ฒฝ์šฐ์—๋Š” ๊ฐ๊ฐ์˜ ๊ฒฝ์šฐ์— ๋Œ€ํ•ด ๊ณ„์‚ฐํ•ด์ค˜์•ผ ํ•˜๋ฏ€๋กœ 5์ œ๊ณฑ๋ฐฐ๊ฐ€ ๋œ๋‹ค.

๋…ผ๋ฌธ์—์„œ๋Š” ์ด ํ™•๋ฅ  ์ค‘์—์„œ ์„ ํƒํ•˜๊ธฐ ์œ„ํ•ด PPO(Proximal Policy Optimiation)์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ๊ฐ์˜ policy๋ฅผ ์„ ์ •ํ•œ๋‹ค.

 

 

 

 


Reference


[1] Zoph et al, Learning Data Augmentation Strategies for Object Detection, 2019

[2] "Understanding Data Augmentation | What is Data Augmentation & how it works?".greatlearningblog, Aug5,2020

Comments