์ฝ˜ํ…์ธ ๋กœ ๊ฑด๋„ˆ๋›ฐ๊ธฐ

์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ์˜ ์ฃผ์š” ๋‹จ๊ณ„ ์ดํ•ดํ•˜๊ธฐ

์†Œ๊ฐœ

Computer vision is a subfield of artificial intelligence (AI) that helps computers see and understand the world like humans do. It processes and analyzes images or videos to extract information, recognize patterns, and make decisions based on that data.



Watch: How to Do [Computer Vision](https://www.ultralytics.com/glossary/computer-vision-cv) Projects | A Step-by-Step Guide

๊ฐ์ฒด ๊ฐ์ง€, ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜, ์ธ์Šคํ„ด์Šค ๋ถ„ํ• ๊ณผ ๊ฐ™์€ ์ปดํ“จํ„ฐ ๋น„์ „ ๊ธฐ์ˆ ์€ ์ž์œจ ์ฃผํ–‰๋ถ€ํ„ฐ ์˜๋ฃŒ ์˜์ƒ์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋‹ค์–‘ํ•œ ์‚ฐ์—… ๋ถ„์•ผ์— ์ ์šฉ๋˜์–ด ๊ฐ€์น˜ ์žˆ๋Š” ์ธ์‚ฌ์ดํŠธ๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์ปดํ“จํ„ฐ ๋น„์ „ ๊ธฐ์ˆ  ๊ฐœ์š”

์ง์ ‘ ์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ๋ฅผ ์ง„ํ–‰ํ•˜๋Š” ๊ฒƒ์€ ์ปดํ“จํ„ฐ ๋น„์ „์— ๋Œ€ํ•ด ์ดํ•ดํ•˜๊ณ  ๋” ๋งŽ์€ ๊ฒƒ์„ ๋ฐฐ์šธ ์ˆ˜ ์žˆ๋Š” ์ข‹์€ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ๋Š” ์—ฌ๋Ÿฌ ๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ ์ฒ˜์Œ์—๋Š” ํ˜ผ๋ž€์Šค๋Ÿฌ์›Œ ๋ณด์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฐ€์ด๋“œ๊ฐ€ ๋๋‚˜๋ฉด ์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ์— ๊ด€๋ จ๋œ ๋‹จ๊ณ„์— ์ต์ˆ™ํ•ด์งˆ ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ”„๋กœ์ ํŠธ์˜ ์‹œ์ž‘๋ถ€ํ„ฐ ๋๊นŒ์ง€ ๋ชจ๋“  ๊ณผ์ •์„ ์‚ดํŽด๋ณด๋ฉด์„œ ๊ฐ ๋‹จ๊ณ„๊ฐ€ ์ค‘์š”ํ•œ ์ด์œ ๋ฅผ ์„ค๋ช…ํ•ด ๋“œ๋ฆฌ๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด์ œ ์‹œ์ž‘ํ•ด์„œ ์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ๋ฅผ ์„ฑ๊ณต์œผ๋กœ ์ด๋Œ์–ด ๋ณด์„ธ์š”!

์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ ๊ฐœ์š”

์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ์— ๊ด€๋ จ๋œ ๊ฐ ๋‹จ๊ณ„์˜ ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ๋…ผ์˜ํ•˜๊ธฐ ์ „์— ์ „์ฒด ํ”„๋กœ์„ธ์Šค๋ฅผ ์‚ดํŽด๋ด…์‹œ๋‹ค. ์˜ค๋Š˜ ์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ๋ฅผ ์‹œ์ž‘ํ•œ๋‹ค๋ฉด ๋‹ค์Œ ๋‹จ๊ณ„๋ฅผ ์ˆ˜ํ–‰ํ•˜๊ฒŒ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค:

  • ํ”„๋กœ์ ํŠธ์˜ ์š”๊ตฌ ์‚ฌํ•ญ์„ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์ด ์ตœ์šฐ์„  ๊ณผ์ œ์ž…๋‹ˆ๋‹ค.
  • ๊ทธ๋Ÿฐ ๋‹ค์Œ ๋ชจ๋ธ ํ•™์Šต์— ๋„์›€์ด ๋˜๋Š” ์ด๋ฏธ์ง€๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ์ •ํ™•ํ•˜๊ฒŒ ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค.
  • ๋‹ค์Œ์œผ๋กœ, ๋ฐ์ดํ„ฐ๋ฅผ ์ •๋ฆฌํ•˜๊ณ  ์ฆ๊ฐ• ๊ธฐ์ˆ ์„ ์ ์šฉํ•˜์—ฌ ๋ชจ๋ธ ํ•™์Šต์„ ์ค€๋น„ํ•ฉ๋‹ˆ๋‹ค.
  • ๋ชจ๋ธ ํ›ˆ๋ จ ํ›„์—๋Š” ๋ชจ๋ธ์„ ์ฒ ์ €ํžˆ ํ…Œ์ŠคํŠธํ•˜๊ณ  ํ‰๊ฐ€ํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์กฐ๊ฑด์—์„œ ์ผ๊ด€๋˜๊ฒŒ ์ž‘๋™ํ•˜๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค.
  • ๋งˆ์ง€๋ง‰์œผ๋กœ ๋ชจ๋ธ์„ ์‹ค์ œ ํ™˜๊ฒฝ์— ๋ฐฐํฌํ•˜๊ณ  ์ƒˆ๋กœ์šด ์ธ์‚ฌ์ดํŠธ์™€ ํ”ผ๋“œ๋ฐฑ์„ ๋ฐ”ํƒ•์œผ๋กœ ์—…๋ฐ์ดํŠธํ•ฉ๋‹ˆ๋‹ค.

์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ ๋‹จ๊ณ„ ๊ฐœ์š”

์ด์ œ ๋ฌด์—‡์„ ๊ธฐ๋Œ€ํ•ด์•ผ ํ•˜๋Š”์ง€ ์•Œ์•˜์œผ๋‹ˆ, ๋ฐ”๋กœ ๋‹จ๊ณ„๋ณ„๋กœ ๋“ค์–ด๊ฐ€์„œ ํ”„๋กœ์ ํŠธ๋ฅผ ์ง„ํ–‰ํ•ด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.

1๋‹จ๊ณ„: ํ”„๋กœ์ ํŠธ์˜ ๋ชฉํ‘œ ์ •์˜ํ•˜๊ธฐ

์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ์˜ ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„๋Š” ํ•ด๊ฒฐํ•˜๊ณ ์ž ํ•˜๋Š” ๋ฌธ์ œ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ์ •์˜ํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค. ์ตœ์ข… ๋ชฉํ‘œ๋ฅผ ์•Œ๋ฉด ์†”๋ฃจ์…˜ ๊ตฌ์ถ•์„ ์‹œ์ž‘ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ํ”„๋กœ์ ํŠธ์˜ ๋ชฉํ‘œ๊ฐ€ ์ง‘์ค‘ํ•ด์•ผ ํ•  ์ปดํ“จํ„ฐ ๋น„์ „ ์ž‘์—…์— ์ง์ ‘์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น˜๊ธฐ ๋•Œ๋ฌธ์— ์ปดํ“จํ„ฐ ๋น„์ „์˜ ๊ฒฝ์šฐ ํŠนํžˆ ๊ทธ๋ ‡์Šต๋‹ˆ๋‹ค.

๋‹ค์Œ์€ ํ”„๋กœ์ ํŠธ ๋ชฉํ‘œ์™€ ์ด๋Ÿฌํ•œ ๋ชฉํ‘œ๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์ปดํ“จํ„ฐ ๋น„์ „ ์ž‘์—…์˜ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ์ž…๋‹ˆ๋‹ค:

  • ๋ชฉํ‘œ: ๊ณ ์†๋„๋กœ์—์„œ ๋‹ค์–‘ํ•œ ์ฐจ๋Ÿ‰ ์œ ํ˜•์˜ ํ๋ฆ„์„ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๊ณ  ๊ด€๋ฆฌํ•˜์—ฌ ๊ตํ†ต ๊ด€๋ฆฌ ๋ฐ ์•ˆ์ „์„ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ•ฉ๋‹ˆ๋‹ค.

    • ์ปดํ“จํ„ฐ ๋น„์ „ ์ž‘์—…: ๋ฌผ์ฒด ๊ฐ์ง€๋Š” ์—ฌ๋Ÿฌ ๋Œ€์˜ ์ฐจ๋Ÿ‰์„ ํšจ์œจ์ ์œผ๋กœ ์ฐพ์•„ ์‹๋ณ„ํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๊ตํ†ต ๋ชจ๋‹ˆํ„ฐ๋ง์— ์ด์ƒ์ ์ž…๋‹ˆ๋‹ค. ์ด ์ž‘์—…์— ๋ถˆํ•„์š”ํ•œ ์„ธ๋ถ€ ์ •๋ณด๋ฅผ ์ œ๊ณตํ•˜๋Š” ์ด๋ฏธ์ง€ ๋ถ„ํ• ๋ณด๋‹ค ๊ณ„์‚ฐ ๋ถ€๋‹ด์ด ์ ๊ธฐ ๋•Œ๋ฌธ์— ๋” ๋น ๋ฅธ ์‹ค์‹œ๊ฐ„ ๋ถ„์„์„ ๋ณด์žฅํ•ฉ๋‹ˆ๋‹ค.
  • ๋ชฉํ‘œ: ์˜๋ฃŒ ์˜์ƒ ์Šค์บ”์—์„œ ์ข…์–‘์˜ ์ •ํ™•ํ•œ ํ”ฝ์…€ ์ˆ˜์ค€ ์œค๊ณฝ์„ ์ œ๊ณตํ•˜์—ฌ ๋ฐฉ์‚ฌ์„ ๊ณผ ์˜์‚ฌ๋ฅผ ๋ณด์กฐํ•˜๋Š” ๋„๊ตฌ๋ฅผ ๊ฐœ๋ฐœํ•ฉ๋‹ˆ๋‹ค.

    • ์ปดํ“จํ„ฐ ๋น„์ „ ์ž‘์—…: ์ด๋ฏธ์ง€ ๋ถ„ํ• ์€ ์ข…์–‘์˜ ํฌ๊ธฐ, ๋ชจ์–‘ ๋ฐ ์น˜๋ฃŒ ๊ณ„ํš์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์ค‘์š”ํ•œ ์ข…์–‘์˜ ์ •ํ™•ํ•˜๊ณ  ์ƒ์„ธํ•œ ๊ฒฝ๊ณ„๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์˜๋ฃŒ ์˜์ƒ์— ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค.
  • ๋ชฉํ‘œ: ์กฐ์ง์˜ ํšจ์œจ์„ฑ๊ณผ ๋ฌธ์„œ ๊ฒ€์ƒ‰์„ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฌธ์„œ(์˜ˆ: ์†ก์žฅ, ์˜์ˆ˜์ฆ, ๋ฒ•๋ฅ  ์„œ๋ฅ˜)๋ฅผ ๋ถ„๋ฅ˜ํ•˜๋Š” ๋””์ง€ํ„ธ ์‹œ์Šคํ…œ์„ ๊ตฌ์ถ•ํ•ฉ๋‹ˆ๋‹ค.

    • Computer Vision Task: Image classification is ideal here as it handles one document at a time, without needing to consider the document's position in the image. This approach simplifies and accelerates the sorting process.

1.5๋‹จ๊ณ„: ์ ํ•ฉํ•œ ๋ชจ๋ธ ๋ฐ ๊ต์œก ์ ‘๊ทผ ๋ฐฉ์‹ ์„ ํƒํ•˜๊ธฐ

ํ”„๋กœ์ ํŠธ ๋ชฉํ‘œ์™€ ์ ํ•ฉํ•œ ์ปดํ“จํ„ฐ ๋น„์ „ ์ž‘์—…์„ ์ดํ•ดํ•œ ํ›„ ํ”„๋กœ์ ํŠธ ๋ชฉํ‘œ๋ฅผ ์ •์˜ํ•˜๋Š” ๋ฐ ์žˆ์–ด ํ•„์ˆ˜์ ์ธ ๋ถ€๋ถ„์€ ์˜ฌ๋ฐ”๋ฅธ ๋ชจ๋ธ๊ณผ ๊ต์œก ์ ‘๊ทผ ๋ฐฉ์‹์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.

๋ชฉํ‘œ์— ๋”ฐ๋ผ ๋จผ์ € ๋ชจ๋ธ์„ ์„ ํƒํ•˜๊ฑฐ๋‚˜ 2๋‹จ๊ณ„์—์„œ ์–ด๋–ค ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ™•์ธํ•œ ํ›„์— ๋ชจ๋ธ์„ ์„ ํƒํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด ํ”„๋กœ์ ํŠธ๊ฐ€ ํŠน์ • ์œ ํ˜•์˜ ๋ฐ์ดํ„ฐ์˜ ๊ฐ€์šฉ์„ฑ์— ํฌ๊ฒŒ ์˜์กดํ•œ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค. ์ด ๊ฒฝ์šฐ ๋ชจ๋ธ์„ ์„ ํƒํ•˜๊ธฐ ์ „์— ๋จผ์ € ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ๋ถ„์„ํ•˜๋Š” ๊ฒƒ์ด ๋” ์‹ค์šฉ์ ์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ˜๋ฉด์— ๋ชจ๋ธ ์š”๊ตฌ ์‚ฌํ•ญ์„ ๋ช…ํ™•ํ•˜๊ฒŒ ์ดํ•ดํ•˜๊ณ  ์žˆ๋‹ค๋ฉด ๋ชจ๋ธ์„ ๋จผ์ € ์„ ํƒํ•œ ๋‹ค์Œ ํ•ด๋‹น ์‚ฌ์–‘์— ๋งž๋Š” ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Choosing between training from scratch or using transfer learning affects how you prepare your data. Training from scratch requires a diverse dataset to build the model's understanding from the ground up. Transfer learning, on the other hand, allows you to use a pre-trained model and adapt it with a smaller, more specific dataset. Also, choosing a specific model to train will determine how you need to prepare your data, such as resizing images or adding annotations, according to the model's specific requirements.

์ฒ˜์Œ๋ถ€ํ„ฐ ๊ต์œกํ•˜๊ธฐ ๋Œ€ ์ „์ด ํ•™์Šต ํ™œ์šฉํ•˜๊ธฐ

Note: When choosing a model, consider its deployment to ensure compatibility and performance. For example, lightweight models are ideal for edge computing due to their efficiency on resource-constrained devices. To learn more about the key points related to defining your project, read our guide on defining your project's goals and selecting the right model.

์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ์˜ ์‹ค์ œ ์ž‘์—…์— ๋“ค์–ด๊ฐ€๊ธฐ ์ „์— ์ด๋Ÿฌํ•œ ์„ธ๋ถ€ ์‚ฌํ•ญ์„ ๋ช…ํ™•ํ•˜๊ฒŒ ์ดํ•ดํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. 2๋‹จ๊ณ„๋กœ ๋„˜์–ด๊ฐ€๊ธฐ ์ „์— ๋‹ค์Œ ์‚ฌํ•ญ์„ ๊ณ ๋ คํ–ˆ๋Š”์ง€ ๋‹ค์‹œ ํ•œ ๋ฒˆ ํ™•์ธํ•˜์„ธ์š”:

  • ํ•ด๊ฒฐํ•˜๋ ค๋Š” ๋ฌธ์ œ๋ฅผ ๋ช…ํ™•ํ•˜๊ฒŒ ์ •์˜ํ•˜์„ธ์š”.
  • ํ”„๋กœ์ ํŠธ์˜ ์ตœ์ข… ๋ชฉํ‘œ๋ฅผ ๊ฒฐ์ •ํ•ฉ๋‹ˆ๋‹ค.
  • ํ•„์š”ํ•œ ํŠน์ • ์ปดํ“จํ„ฐ ๋น„์ „ ์ž‘์—…(์˜ˆ: ๋ฌผ์ฒด ๊ฐ์ง€, ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜, ์ด๋ฏธ์ง€ ๋ถ„ํ• )์„ ์‹๋ณ„ํ•ฉ๋‹ˆ๋‹ค.
  • ๋ชจ๋ธ์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ํ•™์Šตํ• ์ง€, ์•„๋‹ˆ๋ฉด ์ „์ด ํ•™์Šต์„ ์‚ฌ์šฉํ• ์ง€ ๊ฒฐ์ •ํ•˜์„ธ์š”.
  • ์ž‘์—… ๋ฐ ๋ฐฐํฌ ์š”๊ตฌ ์‚ฌํ•ญ์— ์ ํ•ฉํ•œ ๋ชจ๋ธ์„ ์„ ํƒํ•˜์„ธ์š”.

2๋‹จ๊ณ„: ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๋ฐ ๋ฐ์ดํ„ฐ ์–ด๋…ธํ…Œ์ด์…˜

์ปดํ“จํ„ฐ ๋น„์ „ ๋ชจ๋ธ์˜ ํ’ˆ์งˆ์€ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํ’ˆ์งˆ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ์ธํ„ฐ๋„ท์—์„œ ์ด๋ฏธ์ง€๋ฅผ ์ˆ˜์ง‘ํ•˜๊ฑฐ๋‚˜ ์ง์ ‘ ์‚ฌ์ง„์„ ์ฐ๊ฑฐ๋‚˜ ๊ธฐ์กด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์€ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ์„ธํŠธ๋ฅผ ๋‹ค์šด๋กœ๋“œํ•  ์ˆ˜ ์žˆ๋Š” ๋ช‡ ๊ฐ€์ง€ ํ›Œ๋ฅญํ•œ ๋ฆฌ์†Œ์Šค์ž…๋‹ˆ๋‹ค: Google ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ฒ€์ƒ‰ ์—”์ง„, UC ์–ด๋ฐ”์ธ ๋จธ์‹  ๋Ÿฌ๋‹ ๋ฆฌํฌ์ง€ํ† ๋ฆฌ ๋ฐ Kaggle ๋ฐ์ดํ„ฐ ์„ธํŠธ.

Ultralytics ์™€ ๊ฐ™์€ ์ผ๋ถ€ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋Š” ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ์…‹์„ ๊ธฐ๋ณธ์ ์œผ๋กœ ์ง€์›ํ•˜๋ฏ€๋กœ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์‰ฝ๊ฒŒ ์‹œ์ž‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ์—๋Š” ์ธ๊ธฐ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์„ ์›ํ™œํ•˜๊ฒŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์œ ํ‹ธ๋ฆฌํ‹ฐ๊ฐ€ ํฌํ•จ๋˜์–ด ์žˆ์–ด ํ”„๋กœ์ ํŠธ ์ดˆ๊ธฐ ๋‹จ๊ณ„์—์„œ ๋งŽ์€ ์‹œ๊ฐ„๊ณผ ๋…ธ๋ ฅ์„ ์ ˆ์•ฝํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

ํ•˜์ง€๋งŒ ์ด๋ฏธ์ง€๋ฅผ ์ˆ˜์ง‘ํ•˜๊ฑฐ๋‚˜ ์ง์ ‘ ์‚ฌ์ง„์„ ์ฐ๋Š” ๊ฒฝ์šฐ์—๋Š” ๋ฐ์ดํ„ฐ์— ์ฃผ์„์„ ๋‹ฌ์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์ฃผ์„์€ ๋ชจ๋ธ์— ์ง€์‹์„ ์ „๋‹ฌํ•˜๊ธฐ ์œ„ํ•ด ๋ฐ์ดํ„ฐ์— ๋ผ๋ฒจ์„ ๋ถ™์ด๋Š” ํ”„๋กœ์„ธ์Šค์ž…๋‹ˆ๋‹ค. ์ž‘์—…ํ•  ๋ฐ์ดํ„ฐ ์ฃผ์„์˜ ์œ ํ˜•์€ ํŠน์ • ์ปดํ“จํ„ฐ ๋น„์ „ ๊ธฐ์ˆ ์— ๋”ฐ๋ผ ๋‹ค๋ฆ…๋‹ˆ๋‹ค. ๋‹ค์Œ์€ ๋ช‡ ๊ฐ€์ง€ ์˜ˆ์‹œ์ž…๋‹ˆ๋‹ค:

  • ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜: ์ „์ฒด ์ด๋ฏธ์ง€์— ๋‹จ์ผ ํด๋ž˜์Šค๋กœ ๋ ˆ์ด๋ธ”์„ ์ง€์ •ํ•ฉ๋‹ˆ๋‹ค.
  • Object Detection: You'll draw bounding boxes around each object in the image and label each box.
  • Image Segmentation: You'll label each pixel in the image according to the object it belongs to, creating detailed object boundaries.

๋‹ค์–‘ํ•œ ์œ ํ˜•์˜ ์ด๋ฏธ์ง€ ์ฃผ์„

๋ฐ์ดํ„ฐ ์ˆ˜์ง‘๊ณผ ์ฃผ์„ ๋‹ฌ๊ธฐ๋Š” ์‹œ๊ฐ„์ด ๋งŽ์ด ์†Œ์š”๋˜๋Š” ์ˆ˜์ž‘์—…์ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ฃผ์„ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ์ด ๊ณผ์ •์„ ๋” ์‰ฝ๊ฒŒ ์ง„ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์€ ์œ ์šฉํ•œ ๊ฐœ๋ฐฉํ˜• ์ฃผ์„ ๋„๊ตฌ ๋ช‡ ๊ฐ€์ง€์ž…๋‹ˆ๋‹ค: LabeI Studio, CVAT, Labelme.

Step 3: Data Augmentation and Splitting Your Dataset

์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ์ฃผ์„์„ ๋‹จ ํ›„์—๋Š” ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์ „์— ๋จผ์ € ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํ•™์Šต, ๊ฒ€์ฆ ๋ฐ ํ…Œ์ŠคํŠธ ์„ธํŠธ๋กœ ๋ถ„ํ• ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ฆ๊ฐ• ์ „์— ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๋ถ„ํ• ํ•˜๋Š” ๊ฒƒ์€ ๋ณ€๊ฒฝ๋˜์ง€ ์•Š์€ ์›๋ณธ ๋ฐ์ดํ„ฐ์—์„œ ๋ชจ๋ธ์„ ํ…Œ์ŠคํŠธํ•˜๊ณ  ๊ฒ€์ฆํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Š” ๋ชจ๋ธ์ด ๋ณด์ด์ง€ ์•Š๋Š” ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ์— ์–ผ๋งˆ๋‚˜ ์ž˜ ์ผ๋ฐ˜ํ™”๋˜๋Š”์ง€ ์ •ํ™•ํ•˜๊ฒŒ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.

๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„ํ• ํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

  • ํ›ˆ๋ จ ์„ธํŠธ: ๋ชจ๋ธ ํ•™์Šต์— ์‚ฌ์šฉ๋˜๋Š” ๋ฐ์ดํ„ฐ์˜ ๊ฐ€์žฅ ํฐ ๋ถ€๋ถ„์œผ๋กœ, ์ผ๋ฐ˜์ ์œผ๋กœ ์ „์ฒด์˜ 70~80%์— ํ•ด๋‹นํ•ฉ๋‹ˆ๋‹ค.
  • Validation Set: Usually around 10-15% of your data; this set is used to tune hyperparameters and validate the model during training, helping to prevent overfitting.
  • ํ…Œ์ŠคํŠธ ์„ธํŠธ: ๋ฐ์ดํ„ฐ์˜ ๋‚˜๋จธ์ง€ 10~15%๋Š” ํ…Œ์ŠคํŠธ ์„ธํŠธ๋กœ ๋”ฐ๋กœ ๋ณด๊ด€ํ•ฉ๋‹ˆ๋‹ค. ํ•™์Šต์ด ์™„๋ฃŒ๋œ ํ›„ ๋ณด์ด์ง€ ์•Š๋Š” ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.

๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„ํ• ํ•œ ํ›„์—๋Š” ์ด๋ฏธ์ง€ ํšŒ์ „, ํฌ๊ธฐ ์กฐ์ •, ๋’ค์ง‘๊ธฐ ๋“ฑ์˜ ๋ณ€ํ™˜์„ ์ ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ํฌ๊ธฐ๋ฅผ ์ธ์œ„์ ์œผ๋กœ ๋Š˜๋ฆฌ๋Š” ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์„ ์ˆ˜ํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์€ ๋ชจ๋ธ์„ ๋ณ€ํ˜•์— ๋”์šฑ ๊ฒฌ๊ณ ํ•˜๊ฒŒ ๋งŒ๋“ค๊ณ  ๋ณด์ด์ง€ ์•Š๋Š” ์ด๋ฏธ์ง€์— ๋Œ€ํ•œ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ต๋‹ˆ๋‹ค.

๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์˜ ์˜ˆ

Libraries like OpenCV, Albumentations, and TensorFlow offer flexible augmentation functions that you can use. Additionally, some libraries, such as Ultralytics, have built-in augmentation settings directly within its model training function, simplifying the process.

๋ฐ์ดํ„ฐ๋ฅผ ๋” ์ž˜ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด Matplotlib ๋˜๋Š” Seaborn๊ณผ ๊ฐ™์€ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•ด ์ด๋ฏธ์ง€๋ฅผ ์‹œ๊ฐํ™”ํ•˜๊ณ  ๋ถ„ํฌ์™€ ํŠน์„ฑ์„ ๋ถ„์„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ๋ฅผ ์‹œ๊ฐํ™”ํ•˜๋ฉด ํŒจํ„ด, ์ด์ƒ ์ง•ํ›„, ์ฆ๊ฐ• ๊ธฐ์ˆ ์˜ ํšจ๊ณผ๋ฅผ ํŒŒ์•…ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ์‹œ๋งจํ‹ฑ ๊ฒ€์ƒ‰, SQL ์ฟผ๋ฆฌ, ๋ฒกํ„ฐ ์œ ์‚ฌ์„ฑ ๊ฒ€์ƒ‰์„ ํ†ตํ•ด ์ปดํ“จํ„ฐ ๋น„์ „ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ํƒ์ƒ‰ํ•˜๋Š” ๋„๊ตฌ์ธ Ultralytics Explorer๋ฅผ ์‚ฌ์šฉํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.

Ultralytics ํƒ์ƒ‰๊ธฐ ๋„๊ตฌ

๋ฐ์ดํ„ฐ๋ฅผ ์ œ๋Œ€๋กœ ์ดํ•ดํ•˜๊ณ , ๋ถ„ํ• ํ•˜๊ณ , ๋ณด๊ฐ•ํ•˜๋ฉด ์‹ค์ œ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์—์„œ ์ž˜ ์ž‘๋™ํ•˜๋Š” ์ž˜ ํ›ˆ๋ จ๋˜๊ณ , ๊ฒ€์ฆ๋˜๊ณ , ํ…Œ์ŠคํŠธ๋œ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

4๋‹จ๊ณ„: ๋ชจ๋ธ ๊ต์œก

๋ฐ์ดํ„ฐ ์„ธํŠธ๊ฐ€ ํ•™์Šตํ•  ์ค€๋น„๊ฐ€ ๋˜๋ฉด ํ•„์š”ํ•œ ํ™˜๊ฒฝ ์„ค์ •, ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ด€๋ฆฌ, ๋ชจ๋ธ ํ•™์Šต์— ์ง‘์ค‘ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

๋จผ์ € ํ™˜๊ฒฝ์ด ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๊ตฌ์„ฑ๋˜์–ด ์žˆ๋Š”์ง€ ํ™•์ธํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ ์—ฌ๊ธฐ์—๋Š” ๋‹ค์Œ์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค:

  • Installing essential libraries and frameworks like TensorFlow, PyTorch, or Ultralytics.
  • GPU ๋ฅผ ์‚ฌ์šฉํ•˜๋Š” ๊ฒฝ์šฐ CUDA ๋ฐ cuDNN๊ณผ ๊ฐ™์€ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์„ค์น˜ํ•˜๋ฉด GPU ๊ฐ€์†์„ ํ™œ์„ฑํ™”ํ•˜๊ณ  ๊ต์œก ํ”„๋กœ์„ธ์Šค์˜ ์†๋„๋ฅผ ๋†’์ด๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.

Then, you can load your training and validation datasets into your environment. Normalize and preprocess the data through resizing, format conversion, or augmentation. With your model selected, configure the layers and specify hyperparameters. Compile the model by setting the loss function, optimizer, and performance metrics.

Libraries like Ultralytics simplify the training process. You can start training by feeding data into the model with minimal code. These libraries handle weight adjustments, backpropagation, and validation automatically. They also offer tools to monitor progress and adjust hyperparameters easily. After training, save the model and its weights with a few commands.

ํšจ์œจ์ ์ธ ๊ต์œก์„ ์œ„ํ•ด์„œ๋Š” ์ ์ ˆํ•œ ๋ฐ์ดํ„ฐ ์„ธํŠธ ๊ด€๋ฆฌ๊ฐ€ ํ•„์ˆ˜์ ์ด๋ผ๋Š” ์ ์„ ๋ช…์‹ฌํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ ์„ธํŠธ์— ๋ฒ„์ „ ๊ด€๋ฆฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ณ€๊ฒฝ ์‚ฌํ•ญ์„ ์ถ”์ ํ•˜๊ณ  ์žฌํ˜„์„ฑ์„ ๋ณด์žฅํ•˜์„ธ์š”. DVC(๋ฐ์ดํ„ฐ ๋ฒ„์ „ ๊ด€๋ฆฌ)์™€ ๊ฐ™์€ ๋„๊ตฌ๋Š” ๋Œ€๊ทœ๋ชจ ๋ฐ์ดํ„ฐ ์„ธํŠธ๋ฅผ ๊ด€๋ฆฌํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

Step 5: Model Evaluation and Model Finetuning

It's important to assess your model's performance using various metrics and refine it to improve accuracy. Evaluating helps identify areas where the model excels and where it may need improvement. Fine-tuning ensures the model is optimized for the best possible performance.

  • Performance Metrics: Use metrics like accuracy, precision, recall, and F1-score to evaluate your model's performance. These metrics provide insights into how well your model is making predictions.
  • ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ์กฐ์ •: ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์กฐ์ •ํ•˜์—ฌ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ์ตœ์ ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๋“œ ๊ฒ€์ƒ‰ ๋˜๋Š” ๋ฌด์ž‘์œ„ ๊ฒ€์ƒ‰๊ณผ ๊ฐ™์€ ๊ธฐ์ˆ ์„ ์‚ฌ์šฉํ•˜๋ฉด ์ตœ์ ์˜ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ๊ฐ’์„ ์ฐพ๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

  • Fine-Tuning: Make small adjustments to the model architecture or training process to enhance performance. This might involve tweaking learning rates, batch sizes, or other model parameters.

6๋‹จ๊ณ„: ๋ชจ๋ธ ํ…Œ์ŠคํŠธ

์ด ๋‹จ๊ณ„์—์„œ๋Š” ๋ชจ๋ธ์ด ์™„์ „ํžˆ ๋ณด์ด์ง€ ์•Š๋Š” ๋ฐ์ดํ„ฐ์—์„œ๋„ ์ž˜ ์ž‘๋™ํ•˜๋Š”์ง€ ํ™•์ธํ•˜์—ฌ ๋ฐฐํฌํ•  ์ค€๋น„๊ฐ€ ๋˜์—ˆ๋Š”์ง€ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ ํ…Œ์ŠคํŠธ์™€ ๋ชจ๋ธ ํ‰๊ฐ€์˜ ์ฐจ์ด์ ์€ ๋ชจ๋ธ์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ๊ฐœ์„ ํ•˜๋Š” ๊ฒƒ์ด ์•„๋‹ˆ๋ผ ์ตœ์ข… ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๊ฒ€์ฆํ•˜๋Š” ๋ฐ ์ค‘์ ์„ ๋‘”๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค.

๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋Š” ์ผ๋ฐ˜์ ์ธ ๋ฌธ์ œ๋ฅผ ์ฒ ์ €ํžˆ ํ…Œ์ŠคํŠธํ•˜๊ณ  ๋””๋ฒ„๊ทธํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ํ•™์Šต ๋˜๋Š” ์œ ํšจ์„ฑ ๊ฒ€์‚ฌ ์ค‘์— ์‚ฌ์šฉํ•˜์ง€ ์•Š์€ ๋ณ„๋„์˜ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ์„ธํŠธ์—์„œ ๋ชจ๋ธ์„ ํ…Œ์ŠคํŠธํ•˜์„ธ์š”. ์ด ๋ฐ์ดํ„ฐ ์„ธํŠธ๋Š” ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ์ผ๊ด€๋˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•ด ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค๋ฅผ ๋‚˜ํƒ€๋‚ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

Also, address common problems such as overfitting, underfitting, and data leakage. Use techniques like cross-validation and anomaly detection to identify and fix these issues.

Step 7: Model Deployment

๋ชจ๋ธ์„ ์ฒ ์ €ํžˆ ํ…Œ์ŠคํŠธํ–ˆ๋‹ค๋ฉด ์ด์ œ ๋ฐฐํฌํ•  ์ฐจ๋ก€์ž…๋‹ˆ๋‹ค. ๋ฐฐํฌ์—๋Š” ๋ชจ๋ธ์„ ํ”„๋กœ๋•์…˜ ํ™˜๊ฒฝ์—์„œ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ๋งŒ๋“œ๋Š” ์ž‘์—…์ด ํฌํ•จ๋ฉ๋‹ˆ๋‹ค. ์ปดํ“จํ„ฐ ๋น„์ „ ๋ชจ๋ธ์„ ๋ฐฐํฌํ•˜๋Š” ๋‹จ๊ณ„๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:

  • ํ™˜๊ฒฝ ์„ค์ •ํ•˜๊ธฐ: ํด๋ผ์šฐ๋“œ ๊ธฐ๋ฐ˜(AWS, Google Cloud, Azure)์ด๋“  ์—์ง€ ๊ธฐ๋ฐ˜(๋กœ์ปฌ ๋””๋ฐ”์ด์Šค, IoT)์ด๋“  ์„ ํƒํ•œ ๋ฐฐํฌ ์˜ต์…˜์— ํ•„์š”ํ•œ ์ธํ”„๋ผ๋ฅผ ๊ตฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.

  • Exporting the Model: Export your model to the appropriate format (e.g., ONNX, TensorRT, CoreML for YOLO11) to ensure compatibility with your deployment platform.

  • ๋ชจ๋ธ ๋ฐฐํฌํ•˜๊ธฐ: API ๋˜๋Š” ์—”๋“œํฌ์ธํŠธ๋ฅผ ์„ค์ •ํ•˜๊ณ  ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜๊ณผ ํ†ตํ•ฉํ•˜์—ฌ ๋ชจ๋ธ์„ ๋ฐฐํฌํ•ฉ๋‹ˆ๋‹ค.
  • ํ™•์žฅ์„ฑ ๋ณด์žฅ: ๋กœ๋“œ ๋ฐธ๋Ÿฐ์„œ, ์ž๋™ ํ™•์žฅ ๊ทธ๋ฃน, ๋ชจ๋‹ˆํ„ฐ๋ง ๋„๊ตฌ๋ฅผ ๊ตฌํ˜„ํ•˜์—ฌ ๋ฆฌ์†Œ์Šค๋ฅผ ๊ด€๋ฆฌํ•˜๊ณ  ์ฆ๊ฐ€ํ•˜๋Š” ๋ฐ์ดํ„ฐ์™€ ์‚ฌ์šฉ์ž ์š”์ฒญ์„ ์ฒ˜๋ฆฌํ•˜์„ธ์š”.

8๋‹จ๊ณ„: ๋ชจ๋‹ˆํ„ฐ๋ง, ์œ ์ง€ ๊ด€๋ฆฌ ๋ฐ ๋ฌธ์„œํ™”

๋ชจ๋ธ์ด ๋ฐฐํฌ๋œ ํ›„์—๋Š” ์ง€์†์ ์œผ๋กœ ์„ฑ๋Šฅ์„ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๊ณ , ๋ฌธ์ œ๋ฅผ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋„๋ก ์œ ์ง€ ๊ด€๋ฆฌํ•˜๋ฉฐ, ํ–ฅํ›„ ์ฐธ์กฐ ๋ฐ ๊ฐœ์„ ์„ ์œ„ํ•ด ์ „์ฒด ํ”„๋กœ์„ธ์Šค๋ฅผ ๋ฌธ์„œํ™”ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค.

๋ชจ๋‹ˆํ„ฐ๋ง ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ํ•ต์‹ฌ ์„ฑ๊ณผ ์ง€ํ‘œ(KPI)๋ฅผ ์ถ”์ ํ•˜๊ณ  ์ด์ƒ ์ง•ํ›„๋‚˜ ์ •ํ™•๋„ ์ €ํ•˜๋ฅผ ๊ฐ์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋ชจ๋ธ์„ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๋ฉด ์ž…๋ ฅ ๋ฐ์ดํ„ฐ์˜ ๋ณ€๊ฒฝ์œผ๋กœ ์ธํ•ด ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ์ €ํ•˜๋˜๋Š” ๋ชจ๋ธ ๋“œ๋ฆฌํ”„ํŠธ๋ฅผ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์—…๋ฐ์ดํŠธ๋œ ๋ฐ์ดํ„ฐ๋กœ ๋ชจ๋ธ์„ ์ฃผ๊ธฐ์ ์œผ๋กœ ์žฌํ•™์Šตํ•˜์—ฌ ์ •ํ™•๋„์™€ ๊ด€๋ จ์„ฑ์„ ์œ ์ง€ํ•˜์„ธ์š”.

๋ชจ๋ธ ๋ชจ๋‹ˆํ„ฐ๋ง

๋ชจ๋‹ˆํ„ฐ๋ง ๋ฐ ์œ ์ง€ ๊ด€๋ฆฌ ์™ธ์—๋„ ๋ฌธ์„œํ™”๋„ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜, ํ•™์Šต ์ ˆ์ฐจ, ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ, ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ ๋‹จ๊ณ„, ๋ฐฐํฌ ๋ฐ ์œ ์ง€๋ณด์ˆ˜ ์ค‘ ๋ณ€๊ฒฝ ์‚ฌํ•ญ์„ ํฌํ•จํ•œ ์ „์ฒด ํ”„๋กœ์„ธ์Šค๋ฅผ ์ฒ ์ €ํžˆ ๋ฌธ์„œํ™”ํ•˜์„ธ์š”. ๋ฌธ์„œํ™”๊ฐ€ ์ž˜ ๋˜์–ด ์žˆ์œผ๋ฉด ์žฌํ˜„์„ฑ์ด ๋ณด์žฅ๋˜๊ณ  ํ–ฅํ›„ ์—…๋ฐ์ดํŠธ๋‚˜ ๋ฌธ์ œ ํ•ด๊ฒฐ์ด ์‰ฌ์›Œ์ง‘๋‹ˆ๋‹ค. ๋ชจ๋ธ์„ ํšจ๊ณผ์ ์œผ๋กœ ๋ชจ๋‹ˆํ„ฐ๋ง, ์œ ์ง€ ๊ด€๋ฆฌ ๋ฐ ๋ฌธ์„œํ™”ํ•˜๋ฉด ์ˆ˜๋ช… ์ฃผ๊ธฐ ๋™์•ˆ ์ •ํ™•ํ•˜๊ณ  ์‹ ๋ขฐํ•  ์ˆ˜ ์žˆ์œผ๋ฉฐ ๊ด€๋ฆฌํ•˜๊ธฐ ์‰ฌ์šด ์ƒํƒœ๋ฅผ ์œ ์ง€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์ปค๋ฎค๋‹ˆํ‹ฐ์™€ ์†Œํ†ตํ•˜๊ธฐ

์ปดํ“จํ„ฐ ๋น„์ „ ์• ํ˜ธ๊ฐ€ ์ปค๋ฎค๋‹ˆํ‹ฐ์™€ ์—ฐ๊ฒฐํ•˜๋ฉด ์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ๋ฅผ ์ง„ํ–‰ํ•˜๋Š” ๋™์•ˆ ์ง๋ฉดํ•˜๋Š” ๋ชจ๋“  ๋ฌธ์ œ๋ฅผ ์ž์‹  ์žˆ๊ฒŒ ํ•ด๊ฒฐํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ์€ ํšจ๊ณผ์ ์œผ๋กœ ๋ฐฐ์šฐ๊ณ , ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ , ๋„คํŠธ์›Œํฌ๋ฅผ ํ˜•์„ฑํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค.

์ปค๋ฎค๋‹ˆํ‹ฐ ๋ฆฌ์†Œ์Šค

  • GitHub Issues: Check out the YOLO11 GitHub repository and use the Issues tab to ask questions, report bugs, and suggest new features. The active community and maintainers are there to help with specific issues.
  • Ultralytics ๋””์Šค์ฝ”๋“œ ์„œ๋ฒ„: Ultralytics Discord ์„œ๋ฒ„์— ๊ฐ€์ž…ํ•˜์—ฌ ๋‹ค๋ฅธ ์‚ฌ์šฉ์ž ๋ฐ ๊ฐœ๋ฐœ์ž์™€ ์†Œํ†ตํ•˜๊ณ , ์ง€์›์„ ๋ฐ›๊ณ , ์ธ์‚ฌ์ดํŠธ๋ฅผ ๊ณต์œ ํ•˜์„ธ์š”.

๊ณต์‹ ๋ฌธ์„œ

  • Ultralytics YOLO11 Documentation: Explore the official YOLO11 documentation for detailed guides with helpful tips on different computer vision tasks and projects.

์ด๋Ÿฌํ•œ ๋ฆฌ์†Œ์Šค๋ฅผ ์‚ฌ์šฉํ•˜๋ฉด ๋ฌธ์ œ๋ฅผ ๊ทน๋ณตํ•˜๊ณ  ์ปดํ“จํ„ฐ ๋น„์ „ ์ปค๋ฎค๋‹ˆํ‹ฐ์˜ ์ตœ์‹  ํŠธ๋ Œ๋“œ์™€ ๋ชจ๋ฒ” ์‚ฌ๋ก€๋ฅผ ์ตœ์‹  ์ƒํƒœ๋กœ ์œ ์ง€ํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค.

์˜ค๋Š˜ ์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ๋ฅผ ์‹œ์ž‘ํ•˜์„ธ์š”!

์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ์— ๋„์ „ํ•˜๋Š” ๊ฒƒ์€ ํฅ๋ฏธ๋กญ๊ณ  ๋ณด๋žŒ์ฐฌ ์ผ์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด ๊ฐ€์ด๋“œ์˜ ๋‹จ๊ณ„๋ฅผ ๋”ฐ๋ฅด๋ฉด ์„ฑ๊ณต์˜ ํƒ„ํƒ„ํ•œ ํ† ๋Œ€๋ฅผ ๊ตฌ์ถ•ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ฐ ๋‹จ๊ณ„๋Š” ๋ชฉํ‘œ์— ๋ถ€ํ•ฉํ•˜๊ณ  ์‹ค์ œ ์‹œ๋‚˜๋ฆฌ์˜ค์—์„œ ์ž˜ ์ž‘๋™ํ•˜๋Š” ์†”๋ฃจ์…˜์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ๋งค์šฐ ์ค‘์š”ํ•ฉ๋‹ˆ๋‹ค. ๊ฒฝํ—˜์ด ์Œ“์ด๋ฉด์„œ ํ”„๋กœ์ ํŠธ๋ฅผ ๊ฐœ์„ ํ•  ์ˆ˜ ์žˆ๋Š” ๊ณ ๊ธ‰ ๊ธฐ์ˆ ๊ณผ ๋„๊ตฌ๋ฅผ ๋ฐœ๊ฒฌํ•˜๊ฒŒ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ํ˜ธ๊ธฐ์‹ฌ์„ ๊ฐ–๊ณ  ๊ณ„์† ๋ฐฐ์šฐ๋ฉฐ ์ƒˆ๋กœ์šด ๋ฐฉ๋ฒ•๊ณผ ํ˜์‹ ์„ ํƒ๊ตฌํ•˜์„ธ์š”!

์ž์ฃผ ๋ฌป๋Š” ์งˆ๋ฌธ

ํ”„๋กœ์ ํŠธ์— ์ ํ•ฉํ•œ ์ปดํ“จํ„ฐ ๋น„์ „ ์ž‘์—…์„ ์„ ํƒํ•˜๋ ค๋ฉด ์–ด๋–ป๊ฒŒ ํ•ด์•ผ ํ•˜๋‚˜์š”?

์˜ฌ๋ฐ”๋ฅธ ์ปดํ“จํ„ฐ ๋น„์ „ ์ž‘์—…์„ ์„ ํƒํ•˜๋Š” ๊ฒƒ์€ ํ”„๋กœ์ ํŠธ์˜ ์ตœ์ข… ๋ชฉํ‘œ์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง‘๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ๊ตํ†ต ์ƒํ™ฉ์„ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๋ ค๋Š” ๊ฒฝ์šฐ ์—ฌ๋Ÿฌ ์ฐจ๋Ÿ‰ ์œ ํ˜•์„ ์‹ค์‹œ๊ฐ„์œผ๋กœ ์ฐพ์•„ ์‹๋ณ„ํ•  ์ˆ˜ ์žˆ๋Š” ๋ฌผ์ฒด ๊ฐ์ง€๊ฐ€ ์ ํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์˜๋ฃŒ ์˜์ƒ์—์„œ๋Š” ์ด๋ฏธ์ง€ ๋ถ„ํ• ์„ ํ†ตํ•ด ์ข…์–‘์˜ ์„ธ๋ถ€์ ์ธ ๊ฒฝ๊ณ„๋ฅผ ํŒŒ์•…ํ•˜์—ฌ ์ง„๋‹จ ๋ฐ ์น˜๋ฃŒ ๊ณ„ํš์„ ์„ธ์šฐ๋Š” ๋ฐ ์ด์ƒ์ ์ž…๋‹ˆ๋‹ค. ๋ฌผ์ฒด ๊ฐ์ง€, ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜, ์ธ์Šคํ„ด์Šค ๋ถ„ํ• ๊ณผ ๊ฐ™์€ ํŠน์ • ์ž‘์—…์— ๋Œ€ํ•ด ์ž์„ธํžˆ ์•Œ์•„๋ณด์„ธ์š”.

์ปดํ“จํ„ฐ ๋น„์ „ ํ”„๋กœ์ ํŠธ์—์„œ ๋ฐ์ดํ„ฐ ์ฃผ์„์ด ์ค‘์š”ํ•œ ์ด์œ ๋Š” ๋ฌด์—‡์ธ๊ฐ€์š”?

๋ฐ์ดํ„ฐ ์–ด๋…ธํ…Œ์ด์…˜์€ ๋ชจ๋ธ์ด ํŒจํ„ด์„ ์ธ์‹ํ•˜๋„๋ก ๊ฐ€๋ฅด์น˜๋Š” ๋ฐ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค. ์ฃผ์„์˜ ์œ ํ˜•์€ ์ž‘์—…์— ๋”ฐ๋ผ ๋‹ค๋ฆ…๋‹ˆ๋‹ค:

  • ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜: ๋‹จ์ผ ํด๋ž˜์Šค๋กœ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋œ ์ „์ฒด ์ด๋ฏธ์ง€์ž…๋‹ˆ๋‹ค.
  • ๊ฐœ์ฒด ๊ฐ์ง€: ๊ฐ์ฒด ์ฃผ์œ„์— ๊ฒฝ๊ณ„ ์ƒ์ž๋ฅผ ๊ทธ๋ฆฝ๋‹ˆ๋‹ค.
  • ์ด๋ฏธ์ง€ ์„ธ๋ถ„ํ™”: ๊ฐ ํ”ฝ์…€์ด ์†ํ•œ ๊ฐ์ฒด์— ๋”ฐ๋ผ ๋ ˆ์ด๋ธ”์ด ์ง€์ •๋ฉ๋‹ˆ๋‹ค.

๋ผ๋ฒจ ์ŠคํŠœ๋””์˜ค, CVAT, ๋ผ๋ฒจ๋ฏธ์™€ ๊ฐ™์€ ๋„๊ตฌ๊ฐ€ ์ด ๊ณผ์ •์„ ๋„์™€์ค„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘ ๋ฐ ์–ด๋…ธํ…Œ์ด์…˜ ๊ฐ€์ด๋“œ๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.

๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์„ ํšจ๊ณผ์ ์œผ๋กœ ๋ณด๊ฐ•ํ•˜๊ณ  ๋ถ„ํ• ํ•˜๋ ค๋ฉด ์–ด๋–ค ๋‹จ๊ณ„๋ฅผ ๋”ฐ๋ผ์•ผ ํ•˜๋‚˜์š”?

๋ณด๊ฐ• ์ „์— ๋ฐ์ดํ„ฐ ์ง‘ํ•ฉ์„ ๋ถ„ํ• ํ•˜๋ฉด ๋ณ€๊ฒฝ๋˜์ง€ ์•Š์€ ์›๋ณธ ๋ฐ์ดํ„ฐ์—์„œ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๊ฒ€์ฆํ•˜๋Š” ๋ฐ ๋„์›€์ด ๋ฉ๋‹ˆ๋‹ค. ๋‹ค์Œ ๋‹จ๊ณ„๋ฅผ ๋”ฐ๋ฅด์„ธ์š”:

  • ํŠธ๋ ˆ์ด๋‹ ์„ธํŠธ: ๋ฐ์ดํ„ฐ์˜ 70~80%.
  • Validation Set: 10-15% for hyperparameter tuning.
  • ํ…Œ์ŠคํŠธ ์„ธํŠธ: ์ตœ์ข… ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด ๋‚จ์€ 10~15%.

๋ถ„ํ•  ํ›„ ํšŒ์ „, ํฌ๊ธฐ ์กฐ์ •, ๋’ค์ง‘๊ธฐ ๋“ฑ์˜ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ• ๊ธฐ์ˆ ์„ ์ ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ์„ธํŠธ์˜ ๋‹ค์–‘์„ฑ์„ ๋†’์ž…๋‹ˆ๋‹ค. Albumentations ๋ฐ OpenCV์™€ ๊ฐ™์€ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๊ฐ€ ๋„์›€์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. Ultralytics ์—์„œ๋„ ํŽธ์˜๋ฅผ ์œ„ํ•ด ๊ธฐ๋ณธ ์ œ๊ณต ์ฆ๊ฐ• ์„ค์ •์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.

ํ•™์Šต๋œ ์ปดํ“จํ„ฐ ๋น„์ „ ๋ชจ๋ธ์„ ๋ฐฐํฌํ•˜๊ธฐ ์œ„ํ•ด ๋‚ด๋ณด๋‚ด๋ ค๋ฉด ์–ด๋–ป๊ฒŒ ํ•ด์•ผ ํ•˜๋‚˜์š”?

Exporting your model ensures compatibility with different deployment platforms. Ultralytics provides multiple formats, including ONNX, TensorRT, and CoreML. To export your YOLO11 model, follow this guide:

  • ์‚ฌ์šฉ export ํ•จ์ˆ˜์— ์›ํ•˜๋Š” ํ˜•์‹ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ์ถ”๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.
  • ๋‚ด๋ณด๋‚ธ ๋ชจ๋ธ์ด ๋ฐฐํฌ ํ™˜๊ฒฝ(์˜ˆ: ์—์ง€ ๋””๋ฐ”์ด์Šค, ํด๋ผ์šฐ๋“œ)์˜ ์‚ฌ์–‘์— ๋งž๋Š”์ง€ ํ™•์ธํ•ฉ๋‹ˆ๋‹ค.

์ž์„ธํ•œ ๋‚ด์šฉ์€ ๋ชจ๋ธ ๋‚ด๋ณด๋‚ด๊ธฐ ๊ฐ€์ด๋“œ์—์„œ ํ™•์ธํ•˜์„ธ์š”.

๋ฐฐํฌ๋œ ์ปดํ“จํ„ฐ ๋น„์ „ ๋ชจ๋ธ์„ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜๊ณ  ์œ ์ง€ ๊ด€๋ฆฌํ•˜๊ธฐ ์œ„ํ•œ ๋ชจ๋ฒ” ์‚ฌ๋ก€๋Š” ๋ฌด์—‡์ธ๊ฐ€์š”?

๋ชจ๋ธ์˜ ์žฅ๊ธฐ์ ์ธ ์„ฑ๊ณต์„ ์œ„ํ•ด์„œ๋Š” ์ง€์†์ ์ธ ๋ชจ๋‹ˆํ„ฐ๋ง๊ณผ ์œ ์ง€ ๊ด€๋ฆฌ๊ฐ€ ํ•„์ˆ˜์ ์ž…๋‹ˆ๋‹ค. ํ•ต์‹ฌ ์„ฑ๊ณผ ์ง€ํ‘œ(KPI)๋ฅผ ์ถ”์ ํ•˜๊ณ  ์ด์ƒ ์ง•ํ›„๋ฅผ ๊ฐ์ง€ํ•  ์ˆ˜ ์žˆ๋Š” ๋„๊ตฌ๋ฅผ ๊ตฌํ˜„ํ•˜์„ธ์š”. ์—…๋ฐ์ดํŠธ๋œ ๋ฐ์ดํ„ฐ๋กœ ๋ชจ๋ธ์„ ์ •๊ธฐ์ ์œผ๋กœ ์žฌํ•™์Šตํ•˜์—ฌ ๋ชจ๋ธ ํŽธ์ฐจ์— ๋Œ€์‘ํ•˜์„ธ์š”. ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜, ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ๋ฐ ๋ณ€๊ฒฝ ์‚ฌํ•ญ์„ ํฌํ•จํ•œ ์ „์ฒด ํ”„๋กœ์„ธ์Šค๋ฅผ ๋ฌธ์„œํ™”ํ•˜์—ฌ ์žฌํ˜„์„ฑ๊ณผ ํ–ฅํ›„ ์—…๋ฐ์ดํŠธ์˜ ์šฉ์ด์„ฑ์„ ๋ณด์žฅํ•˜์„ธ์š”. ๋ชจ๋‹ˆํ„ฐ๋ง ๋ฐ ์œ ์ง€ ๊ด€๋ฆฌ ๊ฐ€์ด๋“œ์—์„œ ์ž์„ธํžˆ ์•Œ์•„๋ณด์„ธ์š”.


๐Ÿ“… Created 4 months ago โœ๏ธ Updated 2 days ago

๋Œ“๊ธ€