image

paper , code

TL;DR

  • task : object detection, pose estimation, object tracking, label assignment
  • problem : DETR์—์„œ hungarian one-to-one matching์„ ํ•˜๋Š” ๋ถ€๋ถ„ ๋•Œ๋ฌธ์—(๊ทธ ๋•์— NMS ๊ฐ™์€๊ฑธ ์•ˆํ•ด๋„ ๋˜์—ˆ์ง€๋งŒ), positive pair๊ฐ€ ๋„ˆ๋ฌด ์—†์–ด ํ•™์Šต์ด ํšจ์œจ์ ์ด์ง€ ๋ชปํ•จ.
  • idea : hybrid matching. one-to-one matching ํ•˜๋‚˜ ํ•˜๊ณ , one-to-many matching(๊ทธ๋ƒฅ gt ์—ฌ๋Ÿฌ๋ฒˆ ๋ณต์‚ฌํ•˜๋ฉด ๋จ)๋„ ํ•จ. ์ด๊ฑธ auxilary loss์ฒ˜๋Ÿผ ๋ชจ๋“  ๋ ˆ์ด์–ด์— ๋Œ€ํ•ด์„œ ํ•จ.
  • architecture : deformable DETR
  • objective : ๊ฐ task์— ๋งž๋Š” ๋ชฉ์  ํ•จ์ˆ˜
  • baseline : deformable DETR, PETR, 3DETR…
  • data :
  • result : ์„ฑ๋Šฅ gain. ํ•™์Šต ์†๋„๋Š” one-to-one matching์„ ํ• ๋•Œ 1epoch์— 65๋ถ„ ์ •๋„ ์˜€๋‹ค๋ฉด hybrid matching์„ ํ•˜๋ฉด 85๋ถ„
  • contribution : ๊ฐ„๋‹จํ•œ trick์œผ๋กœ ์„ฑ๋Šฅ ๊ฐœ์„ .

Details

hybrid matching์„ ํ•˜๋Š” ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•๋“ค

image

Results

image

image

Group DETR: Fast Training Convergence with Decoupled One-to-Many Label Assignment image

query embedding๋“ค ๋„ฃ์„ ๋•Œ, K๊ฐœ์˜ ๊ทธ๋ฃน์„ ๋‚˜๋ˆ„๊ณ  ๊ทธ ๊ทธ๋ฃน ๋‚ด์—์„œ๋งŒ query๋“ค์ด interaction ํ•  ์ˆ˜ ์žˆ์Œ.