Sign In

DLEBench: Evaluating Small-scale Object Editing Ability for Instruction-based Image Editing Model

์ž‘์„ฑ์ž
  • Haebom
์นดํ…Œ๊ณ ๋ฆฌ
Empty

์ €์ž

Shibo Hong, Boxian Ai, Jun Kuang, Wei Wang, FengJiao Chen, Zhongyuan Peng, Chenhao Huang, Yixin Cao

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ๋…ผ๋ฌธ์€ ๋ช…๋ น์–ด ๊ธฐ๋ฐ˜ ์ด๋ฏธ์ง€ ํŽธ์ง‘ ๋ชจ๋ธ(IIEM)์˜ ์ž‘์€ ๊ฐ์ฒด ํŽธ์ง‘ ๋Šฅ๋ ฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ์ƒˆ๋กœ์šด ๋ฒค์น˜๋งˆํฌ์ธ DLEBench๋ฅผ ์†Œ๊ฐœํ•œ๋‹ค. ๊ธฐ์กด ๋ฒค์น˜๋งˆํฌ๊ฐ€ ์ฃผ๋กœ ์ „์ฒด์ ์ธ ํŽธ์ง‘ ๋Šฅ๋ ฅ์— ์ดˆ์ ์„ ๋งž์ถ”์—ˆ์œผ๋‚˜, DLEBench๋Š” ์ด๋ฏธ์ง€ ์˜์—ญ์˜ 1-10%๋ฅผ ์ฐจ์ง€ํ•˜๋Š” ์ž‘์€ ๊ฐ์ฒด๋ฅผ ๋Œ€์ƒ์œผ๋กœ ๋ณต์žกํ•œ ์ƒํ™ฉ์—์„œ์˜ ํŽธ์ง‘ ๋Šฅ๋ ฅ์„ ์‹ฌ์ธต์ ์œผ๋กœ ํ‰๊ฐ€ํ•œ๋‹ค. ์ œ์•ˆ๋œ ๋ฒค์น˜๋งˆํฌ์™€ ํ‰๊ฐ€ ํ”„๋กœํ† ์ฝœ์€ 10๊ฐœ์˜ IIEM์„ ํ‰๊ฐ€ํ•˜์—ฌ ์ž‘์€ ๊ฐ์ฒด ํŽธ์ง‘ ๋Šฅ๋ ฅ์— ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ๊ฒฉ์ฐจ๊ฐ€ ์กด์žฌํ•จ์„ ๋ฐํ˜€๋‚ด๊ณ , ์ด ๋ถ„์•ผ์˜ ๋ฐœ์ „์„ ์œ„ํ•œ ์ „๋ฌธ์ ์ธ ๋ฒค์น˜๋งˆํฌ์˜ ํ•„์š”์„ฑ์„ ๊ฐ•์กฐํ•œ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
์ •๋ฐ€ํ•œ ๊ตญ์†Œ ํŽธ์ง‘ ๋Šฅ๋ ฅ์˜ ์ค‘์š”์„ฑ ๋ถ€๊ฐ: DLEBench๋Š” ์‹ค์ œ ๋ฐ ์ƒ์„ฑ ์ด๋ฏธ์ง€์—์„œ ์ •๋ฐ€ํ•œ ๊ตญ์†Œ ํŽธ์ง‘ ๋ฐ ์„ธ๋ถ€ ์‚ฌํ•ญ ๊ฐœ์„ ์— ํ•„์ˆ˜์ ์ธ ์ž‘์€ ๊ฐ์ฒด ํŽธ์ง‘ ๋Šฅ๋ ฅ์˜ ์ค‘์š”์„ฑ์„ ๊ฐ•์กฐํ•œ๋‹ค.
โ€ข
์ƒˆ๋กœ์šด ํ‰๊ฐ€ ๊ธฐ์ค€ ๋ฐ ๋ฐฉ๋ฒ•๋ก  ์ œ์‹œ: ์ฃผ๊ด€์„ฑ๊ณผ ๋ชจํ˜ธ์„ฑ์„ ์ตœ์†Œํ™”ํ•˜๋Š” ํ‰๊ฐ€ ํ”„๋กœํ† ์ฝœ๊ณผ ๋„๊ตฌ ์ค‘์‹ฌ ๋ฐ ์˜ค๋ผํด ์•ˆ๋‚ด ๋ชจ๋“œ๋ฅผ ํฌํ•จํ•œ ์ด์ค‘ ๋ชจ๋“œ ํ‰๊ฐ€ ํ”„๋ ˆ์ž„์›Œํฌ๋Š” IIEM์˜ ํ‰๊ฐ€๋ฅผ ๋”์šฑ ๊ฐ๊ด€์ ์ด๊ณ  ์‹ ๋ขฐ์„ฑ ์žˆ๊ฒŒ ๋งŒ๋“ ๋‹ค.
โ€ข
๊ธฐ์กด ๋ชจ๋ธ์˜ ์ž‘์€ ๊ฐ์ฒด ํŽธ์ง‘ ๋Šฅ๋ ฅ ๋ถ€์กฑ: ์‹ค์ œ ์‹คํ—˜ ๊ฒฐ๊ณผ๋Š” ๊ธฐ์กด IIEM๋“ค์ด ์ž‘์€ ๊ฐ์ฒด ํŽธ์ง‘ ๋Šฅ๋ ฅ์—์„œ ์ƒ๋‹นํ•œ ์„ฑ๋Šฅ ๊ฒฉ์ฐจ๋ฅผ ๋ณด์ž„์„ ๋‚˜ํƒ€๋‚ด์–ด, ์ด ๋ถ„์•ผ์— ๋Œ€ํ•œ ์ถ”๊ฐ€์ ์ธ ์—ฐ๊ตฌ ๊ฐœ๋ฐœ์ด ํ•„์š”ํ•จ์„ ์‹œ์‚ฌํ•œ๋‹ค.
โ€ข
ํ•œ๊ณ„์ : ๋ฒค์น˜๋งˆํฌ ์ƒ˜ํ”Œ์˜ ๋‹ค์–‘์„ฑ๊ณผ ๋ณต์žก์„ฑ, ํ‰๊ฐ€ ๊ธฐ์ค€์˜ ์ฃผ๊ด€์„ฑ ์™„์ „ํžˆ ๋ฐฐ์ œ๋Š” ์—ฌ์ „ํžˆ ๋„์ „ ๊ณผ์ œ๋กœ ๋‚จ์•„์žˆ์œผ๋ฉฐ, ํ–ฅํ›„ ๋” ๋งŽ์€ ์‹œ๋‚˜๋ฆฌ์˜ค์™€ ์„ธ๋ถ„ํ™”๋œ ํ‰๊ฐ€ ์ง€ํ‘œ ๊ฐœ๋ฐœ์ด ํ•„์š”ํ•  ์ˆ˜ ์žˆ๋‹ค.
๐Ÿ‘