์ฝ˜ํ…์ธ ๋กœ ๊ฑด๋„ˆ๋›ฐ๊ธฐ

YOLOv4: ๋น ๋ฅด๊ณ  ์ •๋ฐ€ํ•œ ๋ฌผ์ฒด ๊ฐ์ง€

์•Œ๋ ‰์„ธ์ด ๋ณดํ์ฝ”๋ธŒ์Šคํ‚ค( https://github.com/AlexeyAB/darknet)๊ฐ€ 2020๋…„์— ์ถœ์‹œํ•œ ์ตœ์ฒจ๋‹จ ์‹ค์‹œ๊ฐ„ ๋ฌผ์ฒด ๊ฐ์ง€๊ธฐ์ธ YOLOv4์˜ ์„ค๋ช…์„œ ํŽ˜์ด์ง€( Ultralytics )์— ์˜ค์‹  ๊ฒƒ์„ ํ™˜์˜ํ•ฉ๋‹ˆ๋‹ค. YOLOv4๋Š” ์†๋„์™€ ์ •ํ™•์„ฑ ์‚ฌ์ด์—์„œ ์ตœ์ ์˜ ๊ท ํ˜•์„ ์ œ๊ณตํ•˜๋„๋ก ์„ค๊ณ„๋˜์–ด ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ํƒ์›”ํ•œ ์„ ํƒ์ด ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.

YOLOv4 ์•„ํ‚คํ…์ฒ˜ ๋‹ค์ด์–ด๊ทธ๋žจ YOLOv4 ์•„ํ‚คํ…์ฒ˜ ๋‹ค์ด์–ด๊ทธ๋žจ. ๋ฐฑ๋ณธ, ๋„ฅ, ํ—ค๋“œ ๊ตฌ์„ฑ ์š”์†Œ์™€ ์ตœ์ ์˜ ์‹ค์‹œ๊ฐ„ ๊ฐ์ฒด ๊ฐ์ง€๋ฅผ ์œ„ํ•œ ์ƒํ˜ธ ์—ฐ๊ฒฐ๋œ ๋ ˆ์ด์–ด๋ฅผ ํฌํ•จํ•œ YOLOv4์˜ ๋ณต์žกํ•œ ๋„คํŠธ์›Œํฌ ์„ค๊ณ„๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค.

์†Œ๊ฐœ

YOLOv4 stands for You Only Look Once version 4. It is a real-time object detection model developed to address the limitations of previous YOLO versions like YOLOv3 and other object detection models. Unlike other convolutional neural network (CNN) based object detectors, YOLOv4 is not only applicable for recommendation systems but also for standalone process management and human input reduction. Its operation on conventional graphics processing units (GPUs) allows for mass usage at an affordable price, and it is designed to work in real-time on a conventional GPU while requiring only one such GPU for training.

์•„ํ‚คํ…์ฒ˜

YOLOv4๋Š” ์„ฑ๋Šฅ์„ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•ด ํ•จ๊ป˜ ์ž‘๋™ํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ํ˜์‹ ์ ์ธ ๊ธฐ๋Šฅ์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ๊ฐ€์ค‘ ์ž”์—ฌ ์—ฐ๊ฒฐ(WRC), ๊ต์ฐจ ๋‹จ๊ณ„ ๋ถ€๋ถ„ ์—ฐ๊ฒฐ(CSP), ๊ต์ฐจ ๋ฏธ๋‹ˆ ๋ฐฐ์น˜ ์ •๊ทœํ™”(CmBN), ์ž๊ธฐ ์ ๋Œ€ ํ•™์Šต(SAT), ๋ฏธ์‹œ ํ™œ์„ฑํ™”, ๋ชจ์ž์ดํฌ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•, ๋“œ๋กญ๋ธ”๋ก ์ •๊ทœํ™”, CIoU ์†์‹ค ๋“ฑ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ๋Šฅ์„ ๊ฒฐํ•ฉํ•˜์—ฌ ์ตœ์ฒจ๋‹จ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์ผ๋ฐ˜์ ์ธ ๊ฐ์ฒด ๊ฐ์ง€๊ธฐ๋Š” ์ž…๋ ฅ, ๋ฐฑ๋ณธ, ๋ชฉ, ๋จธ๋ฆฌ ๋“ฑ ์—ฌ๋Ÿฌ ๋ถ€๋ถ„์œผ๋กœ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค. YOLOv4์˜ ๋ฐฑ๋ณธ์€ ์ด๋ฏธ์ง€๋„ท์—์„œ ์‚ฌ์ „ ํ•™์Šต๋˜๋ฉฐ ๊ฐ์ฒด์˜ ํด๋ž˜์Šค ๋ฐ ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค. ๋ฐฑ๋ณธ์€ VGG, ResNet, ResNeXt ๋˜๋Š” DenseNet์„ ํฌํ•จํ•œ ์—ฌ๋Ÿฌ ๋ชจ๋ธ์—์„œ ๊ฐ€์ ธ์˜ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํƒ์ง€๊ธฐ์˜ ๋ชฉ ๋ถ€๋ถ„์€ ์—ฌ๋Ÿฌ ๋‹จ๊ณ„์˜ ํŠน์ง• ๋งต์„ ์ˆ˜์ง‘ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋ฉฐ ์ผ๋ฐ˜์ ์œผ๋กœ ์—ฌ๋Ÿฌ ๊ฐœ์˜ ์ƒํ–ฅ์‹ ๊ฒฝ๋กœ์™€ ์—ฌ๋Ÿฌ ๊ฐœ์˜ ํ•˜ํ–ฅ์‹ ๊ฒฝ๋กœ๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ๋จธ๋ฆฌ ๋ถ€๋ถ„์€ ์ตœ์ข… ๊ฐ์ฒด ๊ฐ์ง€ ๋ฐ ๋ถ„๋ฅ˜๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.

๊ณต์งœ ๊ฐ€๋ฐฉ

YOLOv4๋Š” ๋˜ํ•œ ์ถ”๋ก  ๋น„์šฉ์„ ๋Š˜๋ฆฌ์ง€ ์•Š๊ณ ๋„ ํ›ˆ๋ จ ์ค‘์— ๋ชจ๋ธ์˜ ์ •ํ™•๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ธฐ์ˆ ์ธ '๊ณต์งœ ๊ฐ€๋ฐฉ'์ด๋ผ๋Š” ๋ฐฉ๋ฒ•์„ ์‚ฌ์šฉํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์€ ๊ฐ์ฒด ๊ฐ์ง€์— ์‚ฌ์šฉ๋˜๋Š” ์ผ๋ฐ˜์ ์ธ ๊ณต์งœ ๊ฐ€๋ฐฉ ๊ธฐ๋ฒ•์œผ๋กœ, ์ž…๋ ฅ ์ด๋ฏธ์ง€์˜ ๊ฐ€๋ณ€์„ฑ์„ ์ฆ๊ฐ€์‹œ์ผœ ๋ชจ๋ธ์˜ ๊ฒฌ๊ณ ์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์˜ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ๋กœ๋Š” ๊ด‘๋„ ์™œ๊ณก(์ด๋ฏธ์ง€์˜ ๋ฐ๊ธฐ, ๋Œ€๋น„, ์ƒ‰์กฐ, ์ฑ„๋„, ๋…ธ์ด์ฆˆ ์กฐ์ •)๊ณผ ๊ธฐํ•˜ํ•™์  ์™œ๊ณก(์ž„์˜ ํฌ๊ธฐ ์กฐ์ •, ์ž๋ฅด๊ธฐ, ๋’ค์ง‘๊ธฐ, ํšŒ์ „ ์ถ”๊ฐ€)์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ๋ฒ•์€ ๋ชจ๋ธ์ด ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ์ด๋ฏธ์ง€์— ๋” ์ž˜ ์ผ๋ฐ˜ํ™”ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋„์™€์ค๋‹ˆ๋‹ค.

๊ธฐ๋Šฅ ๋ฐ ์„ฑ๋Šฅ

YOLOv4๋Š” ๋ฌผ์ฒด ๊ฐ์ง€์˜ ์†๋„์™€ ์ •ํ™•๋„๋ฅผ ์ตœ์ ํ™”ํ•˜๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค. YOLOv4์˜ ์•„ํ‚คํ…์ฒ˜๋Š” ๋ฐฑ๋ณธ์œผ๋กœ CSPDarknet53, ๋„ฅ์œผ๋กœ PANet, ๊ฐ์ง€ ํ—ค๋“œ๋กœ YOLOv3๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์„ค๊ณ„ ๋•๋ถ„์— YOLOv4๋Š” ์ธ์ƒ์ ์ธ ์†๋„๋กœ ๋ฌผ์ฒด ๊ฐ์ง€๋ฅผ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์–ด ์‹ค์‹œ๊ฐ„ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ YOLOv4๋Š” ์ •ํ™•๋„๊ฐ€ ๋›ฐ์–ด๋‚˜ ๋ฌผ์ฒด ๊ฐ์ง€ ๋ฒค์น˜๋งˆํฌ์—์„œ ์ตœ๊ณ  ์ˆ˜์ค€์˜ ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ–ˆ์Šต๋‹ˆ๋‹ค.

์‚ฌ์šฉ ์˜ˆ

์ด ๊ธ€์„ ์ž‘์„ฑํ•˜๋Š” ์‹œ์ ์„ ๊ธฐ์ค€์œผ๋กœ Ultralytics ์€ ํ˜„์žฌ YOLOv4 ๋ชจ๋ธ์„ ์ง€์›ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ YOLOv4 ์‚ฌ์šฉ์— ๊ด€์‹ฌ์ด ์žˆ๋Š” ์‚ฌ์šฉ์ž๋Š” ์„ค์น˜ ๋ฐ ์‚ฌ์šฉ ์ง€์นจ์„ YOLOv4 GitHub ๋ฆฌํฌ์ง€ํ† ๋ฆฌ์—์„œ ์ง์ ‘ ์ฐธ์กฐํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

๋‹ค์Œ์€ YOLOv4๋ฅผ ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•ด ์ทจํ•  ์ˆ˜ ์žˆ๋Š” ์ผ๋ฐ˜์ ์ธ ๋‹จ๊ณ„์— ๋Œ€ํ•œ ๊ฐ„๋žตํ•œ ๊ฐœ์š”์ž…๋‹ˆ๋‹ค:

  1. YOLOv4 GitHub ๋ฆฌํฌ์ง€ํ† ๋ฆฌ( https://github.com/AlexeyAB/darknet)๋ฅผ ๋ฐฉ๋ฌธํ•˜์„ธ์š”.

  2. README ํŒŒ์ผ์— ์ œ๊ณต๋œ ์ง€์นจ์— ๋”ฐ๋ผ ์„ค์น˜ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฆฌํฌ์ง€ํ† ๋ฆฌ ๋ณต์ œ, ํ•„์š”ํ•œ ์ข…์†์„ฑ ์„ค์น˜, ํ•„์š”ํ•œ ํ™˜๊ฒฝ ๋ณ€์ˆ˜ ์„ค์ •์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.

  3. ์„ค์น˜๊ฐ€ ์™„๋ฃŒ๋˜๋ฉด ๋ฆฌํฌ์ง€ํ† ๋ฆฌ์— ์ œ๊ณต๋œ ์‚ฌ์šฉ ์ง€์นจ์— ๋”ฐ๋ผ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—๋Š” ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฐ์ดํ„ฐ ์„ธํŠธ ์ค€๋น„, ๋ชจ๋ธ ๋งค๊ฐœ๋ณ€์ˆ˜ ๊ตฌ์„ฑ, ๋ชจ๋ธ ํ•™์Šต, ํ•™์Šต๋œ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ๊ฐ์ฒด ๊ฐ์ง€๋ฅผ ์ˆ˜ํ–‰ํ•˜๋Š” ๊ณผ์ •์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค.

๊ตฌ์ฒด์ ์ธ ๋‹จ๊ณ„๋Š” ํŠน์ • ์‚ฌ์šฉ ์‚ฌ๋ก€์™€ YOLOv4 ๋ฆฌํฌ์ง€ํ† ๋ฆฌ์˜ ํ˜„์žฌ ์ƒํƒœ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ๋‹ค๋Š” ์ ์— ์œ ์˜ํ•˜์„ธ์š”. ๋”ฐ๋ผ์„œ YOLOv4 GitHub ๋ฆฌํฌ์ง€ํ† ๋ฆฌ์— ์ œ๊ณต๋œ ์ง€์นจ์„ ์ง์ ‘ ์ฐธ์กฐํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.

์ด๋กœ ์ธํ•ด ๋ถˆํŽธ์„ ๋“œ๋ ค ์ฃ„์†กํ•˜๋ฉฐ, YOLOv4์— ๋Œ€ํ•œ ์ง€์›์ด ๊ตฌํ˜„๋˜๋Š” ๋Œ€๋กœ Ultralytics ์— ์‚ฌ์šฉ ์˜ˆ์‹œ๊ฐ€ ํฌํ•จ๋œ ๋ฌธ์„œ๋ฅผ ์—…๋ฐ์ดํŠธํ•˜๋„๋ก ๋…ธ๋ ฅํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

๊ฒฐ๋ก 

YOLOv4๋Š” ์†๋„์™€ ์ •ํ™•์„ฑ ์‚ฌ์ด์—์„œ ๊ท ํ˜•์„ ์ด๋ฃจ๋Š” ๊ฐ•๋ ฅํ•˜๊ณ  ํšจ์œจ์ ์ธ ๋ฌผ์ฒด ๊ฐ์ง€ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ํ›ˆ๋ จ ์‹œ ๊ณ ์œ ํ•œ ๊ธฐ๋Šฅ๊ณผ ๋‹ค์–‘ํ•œ ๋ฌด๋ฃŒ ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์‹ค์‹œ๊ฐ„ ๋ฌผ์ฒด ๊ฐ์ง€ ์ž‘์—…์—์„œ ํƒ์›”ํ•œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. YOLOv4๋Š” ๊ธฐ์กด์˜ GPU ์„ ํ†ตํ•ด ๋ˆ„๊ตฌ๋‚˜ ํ›ˆ๋ จํ•˜๊ณ  ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์–ด ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์ ‘๊ทผ ๊ฐ€๋Šฅํ•˜๊ณ  ์‹ค์šฉ์ ์ž…๋‹ˆ๋‹ค.

์ธ์šฉ ๋ฐ ๊ฐ์‚ฌ

์‹ค์‹œ๊ฐ„ ๊ฐ์ฒด ๊ฐ์ง€ ๋ถ„์•ผ์—์„œ ํฌ๊ฒŒ ๊ธฐ์—ฌํ•œ YOLOv4 ์ž‘์„ฑ์ž์—๊ฒŒ ๊ฐ์‚ฌ์˜ ๋ง์”€์„ ์ „ํ•ฉ๋‹ˆ๋‹ค:

@misc{bochkovskiy2020yolov4,
      title={YOLOv4: Optimal Speed and Accuracy of Object Detection},
      author={Alexey Bochkovskiy and Chien-Yao Wang and Hong-Yuan Mark Liao},
      year={2020},
      eprint={2004.10934},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

YOLOv4 ๋…ผ๋ฌธ ์›๋ณธ์€ arXiv์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ €์ž๋“ค์€ ์ž์‹ ์˜ ์ž‘์—…์„ ๊ณต๊ฐœํ–ˆ์œผ๋ฉฐ, ์ฝ”๋“œ๋ฒ ์ด์Šค๋Š” GitHub์—์„œ ์•ก์„ธ์Šคํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๋ถ„์•ผ๋ฅผ ๋ฐœ์ „์‹œํ‚ค๊ณ  ๋” ๋งŽ์€ ์ปค๋ฎค๋‹ˆํ‹ฐ๊ฐ€ ์ž์‹ ์˜ ์—ฐ๊ตฌ์— ์ ‘๊ทผํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•œ ์ €์ž๋“ค์˜ ๋…ธ๋ ฅ์— ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค.

์ž์ฃผ ๋ฌป๋Š” ์งˆ๋ฌธ

YOLOv4๋ž€ ๋ฌด์—‡์ด๋ฉฐ ๋ฌผ์ฒด ๊ฐ์ง€์— ์ด ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•ด์•ผ ํ•˜๋Š” ์ด์œ ๋Š” ๋ฌด์—‡์ธ๊ฐ€์š”?

"You Only Look Once ๋ฒ„์ „ 4"์˜ ์•ฝ์ž์ธ YOLOv4๋Š” Alexey Bochkovskiy๊ฐ€ 2020๋…„์— ๊ฐœ๋ฐœํ•œ ์ตœ์ฒจ๋‹จ ์‹ค์‹œ๊ฐ„ ๊ฐ์ฒด ๊ฐ์ง€ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ์†๋„์™€ ์ •ํ™•๋„ ์‚ฌ์ด์—์„œ ์ตœ์ ์˜ ๊ท ํ˜•์„ ์ด๋ฃจ๋ฉฐ ์‹ค์‹œ๊ฐ„ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ๋งค์šฐ ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. YOLOv4์˜ ์•„ํ‚คํ…์ฒ˜์—๋Š” ๊ฐ€์ค‘ ์ž”์—ฌ ์—ฐ๊ฒฐ(WRC), ๊ต์ฐจ ๋‹จ๊ณ„ ๋ถ€๋ถ„ ์—ฐ๊ฒฐ(CSP), ์ž๊ธฐ ์ ๋Œ€์  ํ›ˆ๋ จ(SAT) ๋“ฑ ์—ฌ๋Ÿฌ ํ˜์‹ ์ ์ธ ๊ธฐ๋Šฅ์ด ํ†ตํ•ฉ๋˜์–ด ์žˆ์–ด ์ตœ์ฒจ๋‹จ ๊ฒฐ๊ณผ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ธฐ์กด GPU์—์„œ ํšจ์œจ์ ์œผ๋กœ ์ž‘๋™ํ•˜๋Š” ๊ณ ์„ฑ๋Šฅ ๋ชจ๋ธ์„ ์ฐพ๊ณ  ์žˆ๋‹ค๋ฉด YOLOv4๊ฐ€ ํƒ์›”ํ•œ ์„ ํƒ์ž…๋‹ˆ๋‹ค.

YOLOv4์˜ ์•„ํ‚คํ…์ฒ˜๋Š” ์–ด๋–ป๊ฒŒ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋‚˜์š”?

YOLOv4์˜ ์•„ํ‚คํ…์ฒ˜๋Š” ๋ฐฑ๋ณธ, ๋„ฅ, ํ—ค๋“œ ๋“ฑ ๋ช‡ ๊ฐ€์ง€ ํ•ต์‹ฌ ๊ตฌ์„ฑ ์š”์†Œ๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐฑ๋ณธ์€ ํด๋ž˜์Šค ๋ฐ ๋ฐ”์šด๋”ฉ ๋ฐ•์Šค๋ฅผ ์˜ˆ์ธกํ•˜๋„๋ก ์‚ฌ์ „ ํ•™์Šต๋œ VGG, ResNet ๋˜๋Š” CSPDarknet53๊ณผ ๊ฐ™์€ ๋ชจ๋ธ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชฉ์€ PANet์„ ํ™œ์šฉํ•ด ์—ฌ๋Ÿฌ ๋‹จ๊ณ„์˜ ํŠน์ง• ๋งต์„ ์—ฐ๊ฒฐํ•˜์—ฌ ํฌ๊ด„์ ์ธ ๋ฐ์ดํ„ฐ ์ถ”์ถœ์„ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ YOLOv3์˜ ๊ตฌ์„ฑ์„ ์‚ฌ์šฉํ•˜๋Š” ํ—ค๋“œ๊ฐ€ ์ตœ์ข… ๊ฐ์ฒด ๊ฐ์ง€๋ฅผ ์ˆ˜ํ–‰ํ•ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ YOLOv4๋Š” ๋ชจ์ž์ดํฌ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๋ฐ DropBlock ์ •๊ทœํ™”์™€ ๊ฐ™์€ "๊ณต์งœ ๊ฐ€๋ฐฉ" ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•˜์—ฌ ์†๋„์™€ ์ •ํ™•๋„๋ฅผ ๋”์šฑ ์ตœ์ ํ™”ํ•ฉ๋‹ˆ๋‹ค.

YOLOv4์˜ ๋งฅ๋ฝ์—์„œ '๊ณต์งœ ๊ฐ€๋ฐฉ'์ด๋ž€ ๋ฌด์—‡์ธ๊ฐ€์š”?

"๊ณต์งœ ๋ฐฑ"์ด๋ž€ ์ถ”๋ก  ๋น„์šฉ์„ ๋Š˜๋ฆฌ์ง€ ์•Š๊ณ ๋„ YOLOv4์˜ ํ•™์Šต ์ •ํ™•๋„๋ฅผ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐฉ๋ฒ•์„ ๋งํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ์ˆ ์—๋Š” ๊ด‘๋„ ์™œ๊ณก(๋ฐ๊ธฐ, ๋Œ€๋น„ ์กฐ์ • ๋“ฑ) ๋ฐ ๊ธฐํ•˜ํ•™์  ์™œ๊ณก(ํฌ๊ธฐ ์กฐ์ •, ์ž๋ฅด๊ธฐ, ๋’ค์ง‘๊ธฐ, ํšŒ์ „)๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ํ˜•ํƒœ์˜ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ์ฆ๊ฐ•์€ ์ž…๋ ฅ ์ด๋ฏธ์ง€์˜ ๊ฐ€๋ณ€์„ฑ์„ ์ฆ๊ฐ€์‹œํ‚ด์œผ๋กœ์จ YOLOv4๊ฐ€ ๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ์ด๋ฏธ์ง€์— ๋” ์ž˜ ์ผ๋ฐ˜ํ™”ํ•˜์—ฌ ์‹ค์‹œ๊ฐ„ ์„ฑ๋Šฅ์„ ์ €ํ•˜์‹œํ‚ค์ง€ ์•Š๊ณ  ๊ฒฌ๊ณ ์„ฑ๊ณผ ์ •ํ™•์„ฑ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.

YOLOv4๊ฐ€ ๊ธฐ์กด GPU์˜ ์‹ค์‹œ๊ฐ„ ์˜ค๋ธŒ์ ํŠธ ๊ฐ์ง€์— ์ ํ•ฉํ•˜๋‹ค๊ณ  ์—ฌ๊ฒจ์ง€๋Š” ์ด์œ ๋Š” ๋ฌด์—‡์ธ๊ฐ€์š”?

YOLOv4๋Š” ์†๋„์™€ ์ •ํ™•๋„๋ฅผ ๋ชจ๋‘ ์ตœ์ ํ™”ํ•˜๋„๋ก ์„ค๊ณ„๋˜์–ด ๋น ๋ฅด๊ณ  ์•ˆ์ •์ ์ธ ์„ฑ๋Šฅ์ด ํ•„์š”ํ•œ ์‹ค์‹œ๊ฐ„ ๋ฌผ์ฒด ๊ฐ์ง€ ์ž‘์—…์— ์ด์ƒ์ ์ž…๋‹ˆ๋‹ค. ๊ธฐ์กด GPU์—์„œ ํšจ์œจ์ ์œผ๋กœ ์ž‘๋™ํ•˜๋ฉฐ, ํ•™์Šต๊ณผ ์ถ”๋ก  ๋ชจ๋‘์— ๋‹จ ํ•˜๋‚˜์˜ GPU๋งŒ ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ถ”์ฒœ ์‹œ์Šคํ…œ๋ถ€ํ„ฐ ๋…๋ฆฝํ˜• ํ”„๋กœ์„ธ์Šค ๊ด€๋ฆฌ์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์— ์ ‘๊ทผ ๊ฐ€๋Šฅํ•˜๊ณ  ์‹ค์šฉ์ ์ด๋ฉฐ, ๊ด‘๋ฒ”์œ„ํ•œ ํ•˜๋“œ์›จ์–ด ์„ค์ •์˜ ํ•„์š”์„ฑ์„ ์ค„์—ฌ ์‹ค์‹œ๊ฐ„ ๊ฐ์ฒด ๊ฐ์ง€๋ฅผ ์œ„ํ•œ ๋น„์šฉ ํšจ์œจ์ ์ธ ์†”๋ฃจ์…˜์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Ultralytics ์—์„œ ํ˜„์žฌ ์ง€์›ํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ YOLOv4๋ฅผ ์‹œ์ž‘ํ•˜๋ ค๋ฉด ์–ด๋–ป๊ฒŒ ํ•ด์•ผ ํ•˜๋‚˜์š”?

YOLOv4๋ฅผ ์‹œ์ž‘ํ•˜๋ ค๋ฉด ๊ณต์‹ YOLOv4 GitHub ๋ฆฌํฌ์ง€ํ† ๋ฆฌ๋ฅผ ๋ฐฉ๋ฌธํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ๋ฆฌํฌ์ง€ํ† ๋ฆฌ ๋ณต์ œ, ์ข…์†์„ฑ ์„ค์น˜, ํ™˜๊ฒฝ ๋ณ€์ˆ˜ ์„ค์ • ๋“ฑ์ด ํฌํ•จ๋œ README ํŒŒ์ผ์— ์ œ๊ณต๋œ ์„ค์น˜ ์ง€์นจ์„ ๋”ฐ๋ฅด์„ธ์š”. ์„ค์น˜๊ฐ€ ์™„๋ฃŒ๋˜๋ฉด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์ค€๋น„ํ•˜๊ณ , ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๊ตฌ์„ฑํ•˜๊ณ , ์ œ๊ณต๋œ ์‚ฌ์šฉ ์ง€์นจ์— ๋”ฐ๋ผ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Ultralytics ์€ ํ˜„์žฌ YOLOv4๋ฅผ ์ง€์›ํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ, ๊ฐ€์žฅ ์ตœ์‹ ์˜ ์ž์„ธํ•œ ์ง€์นจ์€ YOLOv4 GitHub๋ฅผ ์ง์ ‘ ์ฐธ์กฐํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.



์ƒ์„ฑ 2023-11-12, ์—…๋ฐ์ดํŠธ 2024-07-04
์ž‘์„ฑ์ž: glenn-jocher (9), sergiuwaxmann (1)

๋Œ“๊ธ€