less than 1 minute read

Reference

SuperPoint: Self-Supervised Interest Point Detection and Description

MagicPoint

The MagicPoint architecture is the SuperPoint architecture without the descriptor head.

MagicPoint๋Š” SuperPoint ์•„ํ‚คํ…์ฒ˜์—์„œ descriptor head๋ฅผ ์ œ์™ธํ•œ ๋ฒ„์ „์œผ๋กœ, ์ฃผ๋กœ ๊ด€์‹ฌ์ (interest point) ๊ฒ€์ถœ์— ์ดˆ์ ์„ ๋งž์ถ˜ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.

MagicPoint์˜ ์ฃผ์š” ํŠน์ง•์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

  • ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ ํ•™์Šต: ๋‹จ์ˆœํ•œ ๊ธฐํ•˜ํ•™์  ๋„ํ˜•๋“ค๋กœ ๊ตฌ์„ฑ๋œ Synthetic Shapes ๋ฐ์ดํ„ฐ์…‹์„ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šต๋ฉ๋‹ˆ๋‹ค.
  • VGG ์Šคํƒ€์ผ ์ธ์ฝ”๋”: ์ด๋ฏธ์ง€์˜ ์ฐจ์›์„ ์ถ•์†Œํ•˜๊ธฐ ์œ„ํ•ด VGG ์Šคํƒ€์ผ์˜ ์ธ์ฝ”๋”๋ฅผ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค.
  • ์ „ํ†ต์  ๋ฐฉ์‹ ๋Œ€๋น„ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ: FAST, Harris, Shi-Tomasi์™€ ๊ฐ™์€ ์ „ํ†ต์ ์ธ keypoint ๊ฒ€์ถœ๊ธฐ๋ณด๋‹ค ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค.
  • ๋…ธ์ด์ฆˆ ๊ฐ•์ธ์„ฑ: ํŠนํžˆ ๋…ธ์ด์ฆˆ๊ฐ€ ์žˆ๋Š” ์ƒํ™ฉ์—์„œ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค.
  • SuperPoint์˜ ๊ธฐ์ดˆ: MagicPoint๋Š” ์‹ค์ œ ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ fine-tuning๊ณผ augmentation supervision์„ ๊ฑฐ์ณ SuperPoint๋กœ ๋ฐœ์ „ํ•ฉ๋‹ˆ๋‹ค.

MagicPoint๋Š” SuperPoint ๊ฐœ๋ฐœ์˜ ์ค‘์š”ํ•œ ์ฒซ ๋‹จ๊ณ„๋กœ, ํ•ฉ์„ฑ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ด ํ•™์Šต๋œ ๊ฐ•๋ ฅํ•œ ๊ด€์‹ฌ์  ๊ฒ€์ถœ๊ธฐ์˜ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค.

image

MagicPoint vs SuperPoint(=MagicPoint + Homographic Adaptation)

In summary, the Homographic Adaptation technique used to transform MagicPoint into SuperPoint gives a large boost in repeatability, especially under large viewpoint changes.

SuperPoint๋Š” ์กฐ๋ช… ๋ณ€ํ™”์— ๊ฐ€์žฅ ๋ฐ˜๋ณต์ ์œผ๋กœ ๋Œ€์‘ํ•˜๋ฉฐ, ์‹œ์  ๋ณ€ํ™”์—์„œ๋„ ๊ฒฝ์Ÿ๋ ฅ์„ ๋ณด์ด๊ณ , ๋ชจ๋“  ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ MagicPoint๋ฅผ ๋Šฅ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.

image

Leave a comment