less than 1 minute read

Reference

ํŠน์ง• ์ถ”์ถœ์„ ์ด์šฉํ•œ ์ด๋ฏธ์ง€ ๋งค์นญ ๊ธฐ์ˆ ์˜ ์ตœ๊ทผ ๋™ํ–ฅ ๋ถ„์„

SuperPoint: Self-Supervised Interest Point Detection and Description

On the other end, LIFT [32], a recently introduced convolutional replacement for SIFT stays close to the traditional patch-based detect then describe recipe. The LIFT pipeline contains interest point detection, orientation estimation and descriptor computation, but additionally requires supervision from a classical SfM system.

LIFT(Learned Invariant Feature Transform)๋Š” ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜์˜ ํŠน์ง•์  ์ถ”์ถœ ๋ฐ ๋งค์นญ ๊ธฐ๋ฒ•์ž…๋‹ˆ๋‹ค.

  • CNN ๊ธฐ๋ฐ˜ ๊ตฌ์กฐ: Convolutional Neural Network(CNN)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํŠน์ง•์  ์ถ”์ถœ, ๋ฐฉํ–ฅ ์ถ”์ •, ๋””์Šคํฌ๋ฆฝํ„ฐ ์ƒ์„ฑ์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค.
  • ์„ธ ๊ฐ€์ง€ ์ฃผ์š” ๊ตฌ์„ฑ ์š”์†Œ:
    • Detector: ํŠน์ง•์  ์œ„์น˜ ๊ฒ€์ถœ
    • Orientation Estimator: ํŠน์ง•์ ์˜ ๋ฐฉํ–ฅ ์ถ”์ •
    • Descriptor: ํŠน์ง•์  ์ฃผ๋ณ€ ์˜์—ญ์˜ ํŠน์„ฑ ๊ธฐ์ˆ 
  • ํ†ตํ•ฉ ํ•™์Šต: ์„ธ ๊ตฌ์„ฑ ์š”์†Œ๋ฅผ ํ•จ๊ป˜ ํ•™์Šต์‹œ์ผœ ์ „์ฒด์ ์ธ ์„ฑ๋Šฅ์„ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค.
  • ๋ฐ์ดํ„ฐ ๊ธฐ๋ฐ˜ ํ•™์Šต: ๋Œ€๋Ÿ‰์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šตํ•˜๋ฏ€๋กœ ๋†’์€ ํŠน์ง•์  ๋งค์นญ ์ •ํ™•๋„์™€ ์ผ๋ฐ˜ํ™” ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.
  • ๊ฐ•์ธ์„ฑ: ์Šค์ผ€์ผ, ํšŒ์ „, ์‹œ์  ๋ณ€ํ™”, ๋…ธ์ด์ฆˆ, ์กฐ๋ช… ๋ณ€ํ™” ๋“ฑ์— ๊ฐ•์ธํ•œ ํŠน์ง•์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค.
  • ํŒจ์น˜ ๊ธฐ๋ฐ˜ ์ฒ˜๋ฆฌ: ์ด๋ฏธ์ง€ ์ „์ฒด๊ฐ€ ์•„๋‹Œ ํŒจ์น˜ ๋‹จ์œ„๋กœ ์ฒ˜๋ฆฌํ•˜์—ฌ ํ•™์Šต์˜ ํšจ์œจ์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค.

LIFT๋Š” ์ „ํ†ต์ ์ธ SIFT์™€ ์œ ์‚ฌํ•œ ์ ‘๊ทผ ๋ฐฉ์‹์„ ์ทจํ•˜์ง€๋งŒ, ๋”ฅ๋Ÿฌ๋‹์„ ํ™œ์šฉํ•˜์—ฌ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋‹ฌ์„ฑํ•˜๊ณ ์ž ํ•ฉ๋‹ˆ๋‹ค. ํŠนํžˆ ๋ทฐํฌ์ธํŠธ ๋ณ€ํ™”๊ฐ€ ํฐ ๊ฒฝ์šฐ์—๋„ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ด๋Š” ๊ฒƒ์ด ํŠน์ง•์ž…๋‹ˆ๋‹ค.

Leave a comment