Skip to content

๐Ÿ›ฐ์ž˜ํ• ๊ฑฐSIA : Satellite Object Detection๐Ÿ›ฐ

Notifications You must be signed in to change notification settings

Chabbbbbo/SIA_TeamProject

Folders and files

NameName
Last commit message
Last commit date

Latest commit

ย 
ย 
ย 
ย 
ย 
ย 
ย 

Repository files navigation

Detection of Aircraft, Vehicles and Ships in Satellite imagery ::
Team ์ž˜ํ• ๊ฑฐSIA

GitHub Visual Studio Code Python PyTorch NumPy Matplotlib Pandas Slack Google Drive



โœ… Detection of Aircraft, Vehicles and Ships in Satellite imagery

๐ŸŒˆย ์ž˜ํ• ๊ฑฐSIAํŒ€ ๊ตฌ์„ฑ ๋ฐ ์—ญํ• 

์ด๋ฆ„ ๊ตฌ์„ฑ ์—ญํ• 
์ฐจ๋ณด๊ฒฝ ํŒ€์žฅ ํ”„๋กœ์ ํŠธ ์ง„ํ–‰ ๋ฐฉํ–ฅ์„ค์ •, Evaluation Metrix ๊ฐœ์„ , ๋ชจ๋ธ ํ•™์Šต ๋ฐ ๋ถ„์„
์ฑ„์ค€๋ณ‘ ํŒ€์› Dataset EDA, ๋ชจ๋ธ ํ•™์Šต ๋ฐ ๋ถ„์„, large image ์ฒ˜๋ฆฌ
ํ•œ์—ฐ๊ทœ ํŒ€์› Dataset EDA, ๋ชจ๋ธ ํ•™์Šต ๋ฐ ๋ถ„์„
์ž„์ƒˆ๋ž€ ํŒ€์› ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ๊ตฌํ˜„(๋ฐ์ดํ„ฐ์…‹ ์ปค์Šคํ„ฐ๋งˆ์ด์ง• ํฌํ•จ ๋ฒ ์ด์Šค๋ผ์ธ ์ฝ”๋“œ์ž‘์„ฑ), ๋ชจ๋ธ ํ•™์Šต ๋ฐ ๋ถ„์„, README ์ž‘์„ฑ
์œคํ˜œ์—ฐ ํŒ€์› ๋ชจ๋ธ ๊ฒฐ๊ณผ ์‹œ๊ฐํ™” (QGIS ๋“ฑ), ๋ชจ๋ธ ํ•™์Šต ๋ฐ ๋ถ„์„

1. ํ”„๋กœ์ ํŠธ ์ฃผ์ œ ๋ฐ ๋ฌธ์ œ์ •์˜

1-1. ํ”„๋กœ์ ํŠธ ์ฃผ์ œ

์œ„์„ฑ ์˜์ƒ์—์„œ์˜ ๊ฐ์ฒด(ํ•ญ๊ณต๊ธฐ, ์„ ๋ฐ•, ์ฐจ๋Ÿ‰) ๋ฅผ ํƒ์ง€ํ•˜๊ธฐ ์œ„ํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๊ฐœ๋ฐœ

์œ„์„ฑ์˜์ƒ์—๋Š” ์ฐจ๋Ÿ‰, ํ•ญ๊ณต๊ธฐ, ์„ ๋ฐ• ๋“ฑ ์•„์ฃผ ๋‹ค์–‘ํ•œ ๊ฐ์ฒด๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ฐ์ฒด๋“ค์„ ์‚ฌ๋žŒ์˜ ๋ˆˆ์œผ๋กœ ์‹ ์†ํ•˜๊ณ  ์ •ํ™•ํ•˜๊ฒŒ ํƒ์ง€ํ•˜๋Š” ๊ฒƒ์€ ๋งค์šฐ ์–ด๋ ค์šด ์ผ์ž…๋‹ˆ๋‹ค.

์ด์— ๋ณธ ํ”„๋กœ์ ํŠธ์—์„œ๋Š” ๋”ฅ๋Ÿฌ๋‹ ๋น„์ „ ๊ธฐ์ˆ ์„ ์ด์šฉํ•˜์—ฌ ์œ„์„ฑ ์˜์ƒ์—์„œ์˜ ๊ฐ์ฒด(ํ•ญ๊ณต๊ธฐ, ์„ ๋ฐ•, ์ฐจ๋Ÿ‰)๋ฅผ ํƒ์ง€ํ•˜๊ธฐ ์œ„ํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ฐœ๋ฐœํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค.

1-2. ํ”„๋กœ์ ํŠธ ๊ณ„ํš

๋ณธ ํ”„๋กœ์ ํŠธ์—์„œ ์ค‘์š”ํ•œ ๊ฒƒ์€ ์œ„์„ฑ์˜์ƒ์ด ๊ฐ€์ง„ ํŠน์ง•์„ ๊ณ ๋ คํ•œ ๊ฐ์ฒดํƒ์ง€ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ตฌ์ถ•ํ•˜์—ฌ์•ผ ํ•œ๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.

์ผ๋ฐ˜์ ์œผ๋กœ ์œ„์„ฑ์˜์ƒ์€ ์ผ๋ฐ˜ ์˜์ƒ์— ๋น„ํ•ด ํšŒ์ „๋œ ๊ฐ์ฒด๊ฐ€ ๋งŽ๋‹ค๋Š” ํŠน์ง•์ด ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ๋‹ค์–‘ํ•œ ๊ฐ๋„๋ฅผ ๊ณ ๋ คํ•œ ๊ฐ์ฒด ํƒ์ง€๋ฅผ ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.

๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๊ฐ์ฒดํƒ์ง€ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์„ฑ๋Šฅ์€ ๋ฐ์ดํ„ฐ์…‹์— ์ง์ ‘์ ์ธ ์˜ํ–ฅ์„ ๋ฐ›๊ธฐ ๋•Œ๋ฌธ์— ๋ฐ์ดํ„ฐ์…‹์˜ ์ž์ฒด์˜ ํŠน์„ฑ(๊ฐ์ฒด๋ถ„ํฌ, ์ด๋ฏธ์ง€ ํฌ๊ธฐ ๋“ฑ)์„ ๊ณ ๋ คํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ตฌ์ถ•ํ•ด์•ผํ•ฉ๋‹ˆ๋‹ค.

๋”ฐ๋ผ์„œ ๋ณธ ํ”„๋กœ์ ํŠธ์—์„œ๋Š” ํฌ๊ฒŒ 1) ๋‹ค์–‘ํ•œ ๊ฐ๋„์—๋„ ๋ฒ”์šฉ์ ์œผ๋กœ ์ ์šฉ์ด ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ์„ ์„ ์ •ํ•˜๋Š” ์ž‘์—…๊ณผ 2) ๋ฐ์ดํ„ฐ์…‹ ์ž์ฒด ํŠน์„ฑ์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ํŠœ๋‹ ํ•˜๋Š” ์ž‘์—… ๋‘๊ฐ€์ง€๋ฅผ ์ง„ํ–‰ํ•ฉ๋‹ˆ๋‹ค.

2. EDA

๋ฐ์ดํ„ฐ ํŠน์„ฑ์„ ๊ณ ๋ คํ•œ ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•ด์„  EDA ๋ฅผ ํ†ตํ•ด ๋ฐ์ดํ„ฐ๊ฐ€ ์–ด๋–ค ํŠน์„ฑ์„ ์ง€๋‹ˆ๊ณ  ์žˆ๋Š”์ง€ ํ™•์ธํ•˜๋Š”๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. EDA๋Š” ์ด๋ฏธ์ง€(Image)์™€ ๊ฐ์ฒด(Object) ๋‘๊ฐ€์ง€ ์ธก๋ฉด์œผ๋กœ ๋‚˜๋ˆ ์„œ ์‚ดํŽด๋ณด๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

2-1. Image

  • ์ด๋ฏธ์ง€ ์‚ฌ์ด์ฆˆ ๋ถ„ํฌ

์ด๋ฏธ์ง€ ์‚ฌ์ด์ฆˆ๊ฐ€ ์•„์ฃผ ๋‹ค์–‘ํ•˜๊ณ , ๋Œ€๋ถ€๋ถ„์˜ ์ด๋ฏธ์ง€์˜ ํ•ด์ƒ๋„๊ฐ€ ํฐํŽธ์ž…๋‹ˆ๋‹ค. ์ด๋ฏธ์ง€ ์‚ฌ์ด์ฆˆ๊ฐ€ ํด ๊ฒฝ์šฐ ํ•™์Šต์— ์–ด๋ ค์›€์ด ์žˆ๊ธฐ์— ์ด๋ฏธ์ง€๋ฅผ ๋ถ„ํ• ํ•˜๋Š”๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.

  • ์ด๋ฏธ์ง€ ์ข…ํšก๋น„๋ณ„ ๋ฐ์ดํ„ฐ์ˆ˜

> ๋Œ€๋ถ€๋ถ„์˜ ์ด๋ฏธ์ง€ ์ข…ํšก๋น„๋Š” ๋Œ€๋ถ€๋ถ„ 1:1, 3:4, 4:3 ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค. 1:1๋กœ Random Cropํ•œ ํ›„ padding ์ ์šฉํ•˜๋Š”๊ฒƒ์ด ์ข‹์•„๋ณด์ž…๋‹ˆ๋‹ค. 

2-2. Object

  • Class๋ณ„ ์ „์ฒด Object ๊ฐฏ์ˆ˜

> Class ๋ณ„๋กœ Object ์ˆ˜๊ฐ€ ์กด์žฌํ•จ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ Vehicle์€ ์ „์ฒด Object์˜ 80%๋ฅผ ์ฐจ์ง€ํ•˜๋Š”๋ฐ, ์ด๋Ÿฌํ•œ ๋ถˆ๊ท ํ˜•์€ ์„ฑ๋Šฅ ์ €ํ•˜์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ์ฃผ์š” ์š”์ธ์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ๋ฐ์ดํ„ฐ ์…‹์—๋Š” ๊ฒ€์ถœํ•ด ๋‚ด๊ณ ์ž ํ•˜๋Š” ํด๋ž˜์Šค (Vehicle, Ship, Airplane) ์™ธ ํด๋ž˜์Šค๋งŒ ์กด์žฌํ•˜๋Š” ๋ฐ์ดํ„ฐ๋„ ์กด์žฌํ•˜๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ๋Š”๋ฐ, ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ๋“ค์„ ํ•™์Šต์‹œํ‚ฌ ๋•Œ ์ œ์™ธ ์ฒ˜๋ฆฌํ• ์ง€, ์œ ์ง€ํ• ์ง€ ๊ณ ๋ฏผํ•ด ๋ณด์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค.
  • 1 Scene ๋‹น Object ๊ฐฏ์ˆ˜ ๋ถ„ํฌ

> ์ด๋ฏธ์ง€ 1์žฅ์— ํฌํ•จ๋œ Object ์ˆ˜๊ฐ€ 1~4๊ฐœ์ธ ์ด๋ฏธ์ง€๊ฐ€ ์ „์ฒด ๋น„์œจ์—์„œ 60%๊ฐ€๋Ÿ‰ ์ฐจ์ง€ํ•จ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. 1Scene ๋‹น ์ตœ๋Œ€ Object๋Š” 1075๊ฐœ๊นŒ์ง€ ์กด์žฌํ•ฉ๋‹ˆ๋‹ค.
  • Class๋ณ„ Object Size ํ†ต๊ณ„

    ํด๋ž˜์Šค ๋‚ด์—์„œ๋„ Object (bounding box) ๋ฉด์ ์ด ์•„์ฃผ ๋‹ค์–‘ํ•˜๊ฒŒ ๋ถ„ํฌํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋ฅผ ํ™•์ธํ•ด ๋ณด๋‹ˆ Labeling Noise ๊ฐ€ ์กด์žฌํ•˜๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

  • Class๋ณ„ Object ์ข…ํšก๋น„ ๋ถ„ํฌ

    Object์˜ size, ์ข…ํšก๋น„ ๋ถ„ํฌ๊ฐ€ ๋„“๊ณ  ํŠนํžˆ Ship์˜ ์ข…ํšก๋น„ ๋ถ„ํฌ๊ฐ€ 0~ 12๊นŒ์ง€ ๋„“๊ฒŒ ๋ถ„ํฌํ•ด์žˆ๋Š” ๊ฒƒ์„ ์•Œ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

    2-3. ๋ฌธ์ œ ํ•ด๊ฒฐ ์ „๋žต

    • Data Imbalance: Baseline model ๋กœ ํ•™์Šต์„ ์ง„ํ–‰ํ•œํ›„ ๊ฒฐ๊ณผ์— ๋”ฐ๋ผ์„œ ์ฒ˜๋ฆฌํ•˜๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.
    • Train input์‹œ, Others ๋งŒ ์กด์žฌํ•˜๋Š” image ์ฒ˜๋ฆฌ : โ€˜Other ํด๋ž˜์Šค๋งŒ ์กด์žฌํ•˜๋Š” ๋ฐ์ดํ„ฐ ์ œ์™ธโ€™ VS โ€˜์œ ์ง€โ€˜๋กœ ๋น„๊ต ์‹คํ—˜์„ ์ง„ํ–‰ํ•˜๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.
    • Scene ๋‹น Object ๋ถ„ํฌ๊ฐ€ ์ œ๊ฐ๊ฐ : Class ๋น„์œจ์— ๋งž์ถฐ ๋ฐ์ดํ„ฐ ์ƒ˜ํ”Œ๋ง์„ ์ง„ํ–‰ํ•˜๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.
    • ์ด๋ฏธ์ง€ ํ•ด์ƒ๋„๊ฐ€ ๋†’์€ ํŽธ (ํ‰๊ท  1000, ์ตœ๋Œ€ 7000): ์ „์ฒ˜๋ฆฌ๋กœ multi split ๋˜๋Š” Random Resize ๋ฅผ ์ง„ํ–‰ํ•˜๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

    3. ๋ฌธ์ œํ•ด๊ฒฐ ๊ณผ์ • 1์ฐจ : Baseline ๋ชจ๋ธ ํ•™์Šต ๋ฐ ํ‰๊ฐ€

    3-1. ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ

    EDA๋ฅผ ํ†ตํ•ด ๋„์ถœํ•ด๋‚ธ ๋ฌธ์ œ์ ๊ณผ ์œ„์„ฑ์˜์ƒ์˜ ํšŒ์ „๋œ ๊ฐ์ฒด ํƒ์ง€๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์„ธ๊ฐ€์ง€ ์ „์ฒ˜๋ฆฌ ๊ณผ์ •์„ ๋ฐ์ดํ„ฐ์…‹์— ์ ์šฉํ•˜์˜€์Šต๋‹ˆ๋‹ค.

    • Resize : ์ด๋ฏธ์ง€ ์‚ฌ์ด์ฆˆ๊ฐ€ ํฌ๋ฉด ํ•™์Šต์— ์–ด๋ ค์›€์ด ์žˆ๊ธฐ๋•Œ๋ฌธ์— EDA๋ฅผ ํ†ตํ•ด ํ™•์ธํ•œ ์ด๋ฏธ์ง€ ์‚ฌ์ด์ฆˆ์™€ ์ข…ํšก๋น„ ๋ถ„ํฌ๋ฅผ ์ฐธ๊ณ ํ•˜์—ฌ ์ด๋ฏธ์ง€ ์‚ฌ์ด์ฆˆ๋ฅผ 1024*1024 ๋กœ ์กฐ์ •ํ•˜์˜€์Šต๋‹ˆ๋‹ค.
    • Crop: ๋‹ค์–‘ํ•œ ****์‚ฌ์ด์ฆˆ๋กœ์˜
    • Flip: ๋‹ค์–‘ํ•œ ๊ฐ๋„์—๋„ ๊ฐ•๊ฑดํ•œ ๋ชจ๋ธ์„ ์œ„ํ•ด์„œ ์ด๋ฏธ์ง€๋ฅผ ๋žœ๋คํ•˜๊ฒŒ 25%์˜ ํ™•๋ฅ ๋กœ '์ˆ˜ํ‰', '์ˆ˜์ง', '๋Œ€๊ฐ์„ ' ๋ฐฉํ–ฅ์œผ๋กœ ํšŒ์ „์„ ์ ์šฉํ•˜์˜€์Šต๋‹ˆ๋‹ค.

    3-2. Baseline Model ์„ ์ •: Oriented Rcnn

    ๊ฐ์ฒดํƒ์ง€ (Object Detection) ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ Localization์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด Feature Map์„ ๊ธฐ๋ฐ˜์œผ๋กœ Object๊ฐ€ ์กด์žฌํ•˜๋Š” ์œ„์น˜์— Bounding Box๋ฅผ ๊ทธ๋ฆฝ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์ธ Bounding Box๋Š” ์ด๋ฏธ์ง€์™€ ์ˆ˜ํ‰์ ์ธ ์ง์‚ฌ๊ฐํ˜•(Horizontal Bounding Box) ํ˜•ํƒœ๋ฅผ ๋งŽ์ด ์‚ฌ์šฉํ•˜๋‚˜, ๊ฐ์ฒด๊ฐ€ ๋งŽ์ด ๋ฐ€์ง‘๋˜์–ด ์žˆ๋Š” ๊ฒฝ์šฐ์—๋Š” Bounding Box๊ฐ€ ๊ฒน์น˜๋Š” ๋ฌธ์ œ๊ฐ€ ์ƒ๊ธฐ๊ธฐ๋„ ํ•ฉ๋‹ˆ๋‹ค.

    ๋”ฐ๋ผ์„œ ๋ณธ ํ”„๋กœ์ ํŠธ์—์„œ๋Š” ์ด๋Ÿฌํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด์„œ Rotated(Oriented) Bounding Box ๋ฅผ ์ง€์›ํ•˜๋Š” ๋ชจ๋ธ ์ค‘ ๋‹จ์ผ ๋ชจ๋ธ ์ค‘ ์„ฑ๋Šฅ์ด ๊ฐ€์žฅ ์šฐ์ˆ˜ํ•œ (Dota ๋ฐ์ดํ„ฐ์…‹ ๊ธฐ์ค€) Oriented RCNN์„ ๊ธฐ๋ณธ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•˜์˜€์Šต๋‹ˆ๋‹ค. ์ƒ์„ธ๊ตฌ์กฐ๋Š” ์•„๋ž˜์™€ ๊ฐ™์Šต๋‹ˆ๋‹ค.

    • backbone : ResNet50 (Pre trained)
    • Neck : feature pyramid network (FPN)
    • Head : RPN head(OrientedRPNHead), RoI Head(RotatedShared2FCBBoxHead, RotatedSingleRoIExtractor)
    • Epoch : 12 (1X)

    3-2. ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ + Baseline Model ํ‰๊ฐ€

Class ๋ณ„ ํ‰๊ฐ€์ง€ํ‘œ(recall, AP, F1Scroe) ๋ฅผ ๋น„๊ตํ•˜๋‹ˆ, EDA๊ณผ์ •์—์„œ ์šฐ๋ คํ–ˆ๋˜ Vehicle์˜ Imbalance ์˜ํ–ฅ์€ ์—†๋Š” ๊ฒƒ์œผ๋กœ ํŒ๋‹จ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.  ์ข€ ๋” ์„ธ๋ฐ€ํ•œ ๋ชจ๋ธ ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด 
Precision ์ง€ํ‘œ๋ฅผ ์ถ”๊ฐ€ํ•ด์ฃผ๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค. 

4. ๋ฌธ์ œํ•ด๊ฒฐ ๊ณผ์ • 2์ฐจ : Other ํด๋ž˜์Šค๋งŒ ์กด์žฌํ•˜๋Š” ๋ฐ์ดํ„ฐ ์ œ์™ธ VS ์œ ์ง€ ์‹คํ—˜

4-1. ์ง€ํ‘œ ํ•ด์„

์ „์ฒ˜๋ฆฌ์™€ ๋ชจ๋ธ์€ ๊ทธ๋Œ€๋กœ ๋ฐ์ดํ„ฐ๋งŒ ๋ฐ”๊ฟ”์„œ ํ•™์Šต์„ ์ง„ํ–‰ํ•˜์˜€๊ณ  ๊ทธ ๊ฒฐ๊ณผ Other ํด๋ž˜์Šค๋งŒ ์กด์žฌํ•˜๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์œ ์ง€ํ•˜๋Š”๊ฒƒ์ด ์ œ์™ธํ•˜๋Š”๊ฒƒ๋ณด๋‹ค mAp,F1Score ๊ฐ’ ๋ชจ๋‘ 0.1 ๊ฐ€๋Ÿ‰ ๋†’์•˜์Šต๋‹ˆ๋‹ค. ๋ฐฐ๊ฒฝ์˜ ๋”ฐ๋ผ์„œ ์•ž์œผ๋กœ์˜ ํ•™์Šต์€ Other ํด๋ž˜์Šค๋งŒ ์กด์žฌํ•˜๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์œ ์ง€ํ•ด์„œ ์ง„ํ–‰ํ•˜๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

์ถ”๊ฐ€์ ์œผ๋กœ Other ํด๋ž˜์Šค๋ฅผ ์œ ์ง€ํ•œ ๋ชจ๋ธ์˜ ๊ฒฐ๊ณผ๋ฅผ ์ƒ์„ธํ•˜๊ฒŒ ํ™•์ธํ•ด๋ณด๋‹ˆ Ship์˜ recall, Precision, ap ๊ฐ’์ด ๋‹ค๋ฅธ Class์—๋น„ํ•ด ์•ฝ 0.1 ๋‚ฎ์€ ๊ฒƒ์„ ํ™•์ธ ํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.

4-2. ํ…Œ์ŠคํŠธ ๊ฒฐ๊ณผ ์‹œ๊ฐํ™”

  • Vehicle

  • Airplane

  • Ship

4-3. ๋ชจ๋ธ ํ‰๊ฐ€ & ๋ฌธ์ œ ํ•ด๊ฒฐ ์ „๋žต ์ˆ˜๋ฆฝ

์ง€ํ‘œ์™€ ์‹œ๊ฐํ™” ๊ฒฐ๊ณผ๋ฅผ ํ™•์ธํ•ด๋ณธ ๊ฒฐ๊ณผ ์ „๋ฐ˜์ ์œผ๋กœ Ship ํด๋ž˜์Šค์—์„œ ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ์•ž์„œ EDA ์—์„œ Ship ํด๋ž˜์Šค์˜ Object ์ข…ํšก๋น„ ๋ถ„ํฌ๊ฐ€ ํƒ€ Class๋ณด๋‹ค ๋„“๊ฒŒ ๋ถ„ํฌ๋˜์–ด ์žˆ๋Š”๊ฒƒ์„ ํ™•์ธํ–ˆ์—ˆ๋Š”๋ฐ, ์ด๋Ÿฌํ•œ ํŠน์ง•์€ Anchor ์ƒ์„ฑ์‹œ ์ข…ํšก๋น„ ์˜ˆ์ธก์ด ์–ด๋ ค์šธ ์ˆ˜ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— ์•„๋ž˜์™€ ๊ฐ™์€ ์ „๋žต์„ ์„ธ์› ์Šต๋‹ˆ๋‹ค.

  • Anchor generator์˜ Ratio, Size ์กฐ์ ˆ
  • Anchor Free model (Rotated FCOS)์ง„ํ–‰

5. ๋ฌธ์ œํ•ด๊ฒฐ ๊ณผ์ • 3์ฐจ : Anchor Free ๋ชจ๋ธ ํ•™์Šต ๋ฐ ํ‰๊ฐ€

5-1. ********Anchor Free ๋ชจ๋ธ ์„ ์ •

Oriented FCOS model

5-1. Anchor Free ๋ชจ๋ธ ํ‰๊ฐ€

1 Stage Model ํŠน์„ฑ์ƒ ๊ธฐ์กด ๋ชจ๋ธ์— ๋น„ํ•ด Evaluation ๊ฐ’์ด ์ „์ฒด์ ์œผ๋กœ ๋–จ์–ด์ง€๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Anchor์˜ ์˜ํ–ฅ์œผ๋กœ Ship์˜ ๊ฒฐ๊ณผ๊ฐ’์ด ๋–จ์–ด์ง„๋‹ค๋ฉด Anchor free ๋ชจ๋ธ์—์„  ๋น„์Šทํ•œ ๊ฐ’์ด ๋‚˜์˜ค๊ธธ ์›ํ–ˆ์œผ๋‚˜, ์—ฌ๊ธฐ์„œ๋„ Ship์ด ํŠนํžˆ ๋‚ฎ์€๊ฒƒ์„ ํ™•์ธ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ํ•ด๋‹น ๋ชจ๋ธ์€ ์ฑ„ํƒ์„ ํ•˜์ง€ ์•Š๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.

6. ๋ฌธ์ œํ•ด๊ฒฐ ๊ณผ์ • 4์ฐจ : ๊ธฐ์กด ๋ชจ๋ธ์˜ Anchor Generator ๋ณด์™„

6-1. Oriented Rcnn ๋ชจ๋ธ ํŠœ๋‹

Anchor Generator ์˜ scale๊ณผ ratio ๋ฅผ ๋‹ค์–‘ํ•˜๊ฒŒ ๋ณ€๊ฒฝํ•˜๋ฉด์„œ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํŒŒ์•…ํ•˜์˜€์Šต๋‹ˆ๋‹ค.

ratio ๋ฅผ ์„ ์ •ํ•˜๋Š” ๊ธฐ์ค€์€ EDA์—์„œ ์‚ดํŽด๋ดค๋˜ Ship ํด๋ž˜์Šค์˜ Object ์ข…ํšก๋น„ ๋ถ„ํฌ๋ฅผ ์ฐธ๊ณ ํ•˜์—ฌ [3,4,5] ๋กœ ์„ ์ •์„ ํ•˜์˜€์Šต๋‹ˆ๋‹ค.

6-2. Oriented Rcnn ๋ชจ๋ธ ํŠœ๋‹ ๊ฒฐ๊ณผ ํ•ด์„

  • Ratio์— ๋”ฐ๋ฅธ ๊ฒฐ๊ณผ ํ‰๊ฐ€

  • Ratio์— ๋”ฐ๋ฅธ ๊ฒฐ๊ณผ ์‹œ๊ฐํ™”

  • Size์— ๋”ฐ๋ฅธ Ship ์ง€ํ‘œ ๋น„๊ต

  • ๊ธฐ์กด ๋ชจ๋ธ๊ณผ ํŠœ๋‹ํ•œ ๋ชจ๋ธ Evaluation ๋น„๊ต

  • Size์— ๋”ฐ๋ฅธ ์‹œ๊ฐํ™”

7. ๋ฌธ์ œํ•ด๊ฒฐ ๊ณผ์ • 5์ฐจ : Class Imbalance ํ•ด๊ฒฐ

EDA์—์„œ ํ™•์ธํ•œ Class Imbalance ์—๋„ ์ž˜ ์ž‘๋™ํ•˜๋Š” ๋ชจ๋ธ์„ ๊ตฌ์ถ•ํ•˜๊ณ ์ž ๋” ๋ณด์™„ํ•˜๊ณ ์ž ์ƒ˜ํ”Œ๋ง์„ ์ง„ํ–‰ํ•˜์˜€์Šต๋‹ค.

  • Scene ๋‹น Object ๋ถ„ํฌ๊ฐ€ ์ œ๊ฐ๊ฐ : Class ๋น„์œจ์— ๋งž์ถฐ ๋ฐ์ดํ„ฐ ์ƒ˜ํ”Œ๋ง์„ ์ง„ํ–‰ํ•˜๋„๋ก ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค

8. ํ”„๋กœ์ ํŠธ ํšŒ๊ณ 

์ด๋ฆ„ ์†Œ๊ฐ
์ฐจ๋ณด๊ฒฝ (ํŒ€์žฅ) anchor-based detector๋Š” class์˜ ์ข…ํšก๋น„์˜ ๋ถ„ํฌ๊ฐ€ ๋‹ค๋ฆ„์— ๋”ฐ๋ผ detection ์ •ํ™•๋„์˜ ํŽธ์ฐจ๊ฐ€ ํฐ ํŽธ์ธ ๊ฒƒ์œผ๋กœ ํ™•์ธํ•  ์ˆ˜ ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋Ÿฌ๋ฏ€๋กœ 3๊ฐ€์ง€ class๋ฅผ ๋ชจ๋‘ ํƒ์ง€ํ•˜๋Š” ๋ชจ๋ธ์„ ๋งŒ๋“ค๊ธฐ๋ณด๋‹ค๋Š” vehicle, airplane, ship์„ ๊ฐ๊ฐ detection ํ•˜๋Š” ๋ชจ๋ธ์„ ๋งŒ๋“ค๊ณ  ์ด๋ฅผ ์•™์ƒ๋ธ” ํ•˜๋Š” ๊ธฐ๋ฒ•์„ ์ ์šฉํ•˜๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค.
์ฑ„์ค€๋ณ‘ (ํŒ€์›) ์ธ๊ณต์œ„์„ฑ ์ด๋ฏธ์ง€์˜ ํŠน์„ฑ์ƒ ํ•™์Šต์— ๋งŽ์€ ์ž์›์ด ํ•„์š”ํ•˜๊ณ  ๊ฐ€์šฉํ•  ์ˆ˜ ์žˆ๋Š” vram์ด ์ž‘์•„ ์•„์‰ฌ์› ์Šต๋‹ˆ๋‹ค. ์ฐจํ›„ ์ด๋ฏธ์ง€์˜ patch, split์„ ๋‹ค์–‘ํ•˜๊ฒŒ ์ ์šฉํ•˜์—ฌ ์›๋ณธ๊ณผ์˜ ๊ฒฐ๊ด๊ฐ’ ์ฐจ์ด๋ฅผ ํ™•์ธํ•ด๋ณด๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค.
ํ•œ์—ฐ๊ทœ (ํŒ€์›) ๊ฐ class๋ณ„๋กœ ์ ํ•ฉํ•œ ๋ชจ๋ธ์„ ์ ์šฉํ•˜๋Š” ์•™์ƒ๋ธ” ๊ธฐ๋ฒ•์„ ํ•ด๋ณด์ง€ ๋ชปํ•ด ์•„์‰ฌ์› ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ, ์œ„์„ฑ ์˜์ƒ ํŠน์„ฑ์— ๋งž๊ฒŒ ๋ชจ๋ธ์„ ์ ์šฉํ•˜๋Š” ๋ฒ•๊ณผ ๋ชจ๋ธ ๊ฐœ์„  ๊ณผ์ •์œผ๋กœ ๋…ผ๋ฆฌ์ ์œผ๋กœ ํ•ด๊ฒฐํ•ด๋ณด๋Š” ์ข‹์€ ๊ฒฝํ—˜์ด์˜€์Šต๋‹ˆ๋‹ค.
์ž„์ƒˆ๋ž€ (ํŒ€์›) ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์—์„œ ์ œ๊ณตํ•˜๋Š” ์ฝ”๋“œ ์™ธ์— ์šฐ๋ฆฌ์—๊ฒŒ ํ•„์š”ํ•œ ์ฝ”๋“œ๋ฅผ ์ถ”๊ฐ€ํ•˜๊ณ  ์ ์šฉํ•˜๋Š” ๊ฒฝํ—˜์„ ํ•  ์ˆ˜ ์žˆ์–ด ํฅ๋ฏธ๋กœ์› ์Šต๋‹ˆ๋‹ค. ํ”„๋กœ์ ํŠธ๋ฅผ ์ง„ํ–‰ํ•˜๋ฉฐ ํŒ€์›๊ฐ„ ์ฝ”๋“œ์˜ ํ†ต์ผ๊ณผ ๊ณต์œ ๊ฐ€ ์ค‘์š”ํ•˜๋‹ค๋Š” ์ ์„ ๋Š๊ผˆ์Šต๋‹ˆ๋‹ค. ๋˜ํ•œ ์‹œ๊ฐ„์ด ๋„ˆ๋ฌด ์—†์–ด ๋” ๋งŽ์€ ์‹œ๋„๋ฅผ ํ•˜์ง€ ๋ชปํ•œ์ ์ด ๋งŽ์ด ์•„์‰ฝ์Šต๋‹ˆ๋‹ค.
์œคํ˜œ์—ฐ (ํŒ€์›) mAP์™€ ๊ฐ™์€ ์ˆ˜์น˜์ƒ์˜ ๋ณ€ํ™”๋ณด๋‹ค๋Š” ์‹œ๊ฐํ™”๋ฅผ ํ†ตํ•ด ํ›ˆ๋ จํ•œ ๋ชจ๋ธ์ด ๊ฐ์ฒด์˜ size๋‚˜ ratio์— ๋”ฐ๋ฅธ ์˜ํ–ฅ์„ ์ž˜ ์บ์น˜ํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ™•์ธํ•˜๋Š” ๊ณผ์ •์ด ๋งค์šฐ ์ค‘์š”ํ•˜๋‹ค๋Š” ์ ์„ ๋Š๊ผˆ์Šต๋‹ˆ๋‹ค.

About

๐Ÿ›ฐ์ž˜ํ• ๊ฑฐSIA : Satellite Object Detection๐Ÿ›ฐ

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 4

  •  
  •  
  •  
  •