1 minute read

Reference

FRSR: Framework for real-time scene reconstruction in robot-assistedminimally invasive surgery

We did not downsample the disparity channels after forming the cost volume to keep the larger disparity hypothesis range.

Cost Volume:

Stereo matching์ด๋‚˜ optical flow์™€ ๊ฐ™์€ 3D ๋น„์ „ ์ž‘์—…์—์„œ ๋‘ ์ด๋ฏธ์ง€ ๊ฐ„์˜ ๋น„์šฉ(cost)์„ ๊ณ„์‚ฐํ•˜์—ฌ ๊ฐ ํ”ฝ์…€์˜ ๋งค์นญ ๊ฐ€๋Šฅ์„ฑ์„ ํ‘œํ˜„ํ•˜๋Š” 3์ฐจ์› ๋ฐ์ดํ„ฐ ๊ตฌ์กฐ์ž…๋‹ˆ๋‹ค.

์ผ๋ฐ˜์ ์œผ๋กœ, cost volume์˜ channel ์ˆ˜๋Š” ๊ฐ disparity ํ›„๋ณด์— ๋Œ€ํ•œ ๋น„์šฉ์„ ์ €์žฅํ•ฉ๋‹ˆ๋‹ค.

๋”ฐ๋ผ์„œ cost volume์˜ channel์€ disparity range์™€ ๊ด€๋ จ์ด ์žˆ์Šต๋‹ˆ๋‹ค.

Disparity Hypothesis:

Disparity hypothesis๋Š” ๊ฐ ํ”ฝ์…€์—์„œ ๊ฐ€์งˆ ์ˆ˜ ์žˆ๋Š” disparity์˜ ํ›„๋ณด ๊ฐ’์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.

์˜ˆ๋ฅผ ๋“ค์–ด, disparity range๊ฐ€ [0, 64]๋ผ๋ฉด ๊ฐ ํ”ฝ์…€์€ 65๊ฐœ์˜ disparity hypothesis๋ฅผ ๊ฐ€์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Cost Volume์—์„œ์˜ Channel ์ˆ˜:

  • ์ผ๋ฐ˜์ ์œผ๋กœ disparity hypothesis์˜ ์ˆ˜์™€ ๋™์ผํ•ฉ๋‹ˆ๋‹ค.
  • ํ•˜์ง€๋งŒ channel์€ hypothesis ์ž์ฒด๋ฅผ ์˜๋ฏธํ•˜์ง€ ์•Š๊ณ , ๊ฐ hypothesis์— ๋Œ€ํ•œ cost ๊ฐ’์„ ์ €์žฅํ•ฉ๋‹ˆ๋‹ค.
  • Disparity Hypothesis๋Š” ๋‹จ์ˆœํžˆ disparity ๊ฐ’์˜ ํ›„๋ณด(range)๋ฅผ ๊ฐ€๋ฆฌํ‚ค๋Š” ๊ฐœ๋…์ด๋ฉฐ, cost volume์˜ channel ์ˆ˜์™€ ํ˜ผ๋™ํ•ด์„œ๋Š” ์•ˆ ๋ฉ๋‹ˆ๋‹ค.

Cost Volume์˜ ์˜๋ฏธ์™€ ์—ญํ• 

On the other hand, we could also consider enhancing the fine feature extraction ability of the model in a complex surgical environment, and a possible solution is to concatenate special layers focusing on extracting local fine features when constructing the cost volume, although it will slow down the prediction speed of the model.

์—ฌ๊ธฐ์„œ cost volume์€ 3D ๋ณต์› ๋˜๋Š” ๋งค์นญ ์ž‘์—…์—์„œ ๊ฐ ํ”ฝ์…€์ด๋‚˜ ํฌ์ธํŠธ์˜ ํŠน์ • ๊ฐ€์„ค(disparity hypothesis ๋˜๋Š” depth hypothesis) ๊ฐ„์˜ ๋น„์šฉ์„ ํ‘œํ˜„ํ•œ 3์ฐจ์› ๋ฐ์ดํ„ฐ ๊ตฌ์กฐ๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.

์ด ๋ฌธ๋งฅ์—์„œ๋Š” ์ˆ˜์ˆ  ํ™˜๊ฒฝ์—์„œ ๋ณต์žกํ•œ ๋ฌผ์ฒด(์˜ˆ: ์–‡์€ ์ˆ˜์ˆ ์šฉ ์‹ค ๋“ฑ)๋ฅผ ์ •ํ™•ํžˆ ์žฌ๊ตฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด cost volume์˜ ์—ญํ• ์ด ๊ฐ•์กฐ๋ฉ๋‹ˆ๋‹ค.

  • Cost volume์€ ํ”ฝ์…€ ๋˜๋Š” ํฌ์ธํŠธ ๊ฐ„์˜ ์ƒ๊ด€๊ด€๊ณ„(similarity)๋ฅผ ๊ณ„์‚ฐํ•˜์—ฌ ์ €์žฅํ•ฉ๋‹ˆ๋‹ค.
  • ๊ฐ ์œ„์น˜์—์„œ ๋‹ค์–‘ํ•œ ๊นŠ์ด ๊ฐ€์„ค(disparity hypothesis)์— ๋Œ€ํ•œ ๋น„์šฉ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ, ์ด ๋น„์šฉ์ด ๋‚ฎ์„์ˆ˜๋ก ๋งค์นญ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค.
  • ์ˆ˜์ˆ ์šฉ ์‹ค(surgical suture)๊ณผ ๊ฐ™์€ ์ž‘๊ณ  ์„ฌ์„ธํ•œ ๊ตฌ์กฐ๋ฌผ์„ ๋ฐฐ๊ฒฝ์—์„œ ๋ถ„๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด, cost volume์ด ์ค‘์š”ํ•œ ์—ญํ• ์„ ํ•ฉ๋‹ˆ๋‹ค.
  • ํ•˜์ง€๋งŒ ๊ธฐ์กด์˜ ๋ชจ๋ธ์€ ์ด๋Ÿฌํ•œ ์„ธ๋ฐ€ํ•œ(local fine) ํŠน์ง•์„ ์ถฉ๋ถ„ํžˆ ์ถ”์ถœํ•˜์ง€ ๋ชปํ•˜๊ธฐ ๋•Œ๋ฌธ์—, ๋ณต์žกํ•œ ์ˆ˜์ˆ  ํ™˜๊ฒฝ์—์„œ ์ผ๋ถ€ ๋ฌผ์ฒด๋ฅผ ์ •ํ™•ํžˆ ์žฌ๊ตฌ์„ฑํ•˜์ง€ ๋ชปํ•˜๋Š” ํ•œ๊ณ„๊ฐ€ ๋“œ๋Ÿฌ๋‚ฌ์Šต๋‹ˆ๋‹ค.
    • ํ•ด๊ฒฐ์ฑ…์œผ๋กœ์„œ cost volume ๊ฐ•ํ™”:
      • ์ด ๋ฌธ๋งฅ์—์„œ ์ œ์•ˆ๋œ ํ•ด๊ฒฐ์ฑ…์€ local fine feature๋ฅผ ์ถ”์ถœํ•  ์ˆ˜ ์žˆ๋Š” layers๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ cost volume์„ ๊ฐ•ํ™”ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
      • ์ฆ‰, ๋ณต์žกํ•œ ํ™˜๊ฒฝ์—์„œ๋„ ์ž‘์€ ๊ตฌ์กฐ๋ฌผ์„ ๊ตฌ๋ณ„ํ•  ์ˆ˜ ์žˆ๋„๋ก cost volume์— ์ •๊ตํ•œ ํŠน์ง• ์ถ”์ถœ ๋Šฅ๋ ฅ์„ ๋ถ€์—ฌํ•˜์ž๋Š” ์ ‘๊ทผ์ž…๋‹ˆ๋‹ค.

Leave a comment