Giter Club home page Giter Club logo

level2-data-annotation_cv-level2-cv-11's Introduction

๐ŸŽจ [Pstage] CV 11์กฐ CanVas

แ„‰แ…ณแ„แ…ณแ„…แ…ตแ†ซแ„‰แ…ฃแ†บ 2022-04-26 แ„‹แ…ฉแ„’แ…ฎ 10 41 27

  • ๋Œ€ํšŒ ๊ธฐ๊ฐ„ : 2022.04.14 ~ 2022.04.21
  • ๋ชฉ์  : ์ด๋ฏธ์ง€ ๋‚ด ๊ธ€์ž ์˜์—ญ ๊ฒ€์ถœ(Text Detection)

Overview

์นด๋ฉ”๋ผ๋กœ ์นด๋“œ ๋‚ด ์นด๋“œ ๋ฒˆํ˜ธ ์ž๋™ ์ธ์‹, ์ฐจ๋Ÿ‰ ๋ฒˆํ˜ธ ์ž๋™ ์ธ์‹ ๋“ฑ OCR (Optimal Character Recognition) ๊ธฐ์ˆ ์€ ์‚ฌ๋žŒ์ด ์ง์ ‘ ์“ฐ๊ฑฐ๋‚˜ ์ด๋ฏธ์ง€ ์†์— ์žˆ๋Š” ๋ฌธ์ž๋ฅผ ์–ป์€ ๋‹ค์Œ ์ด๋ฅผ ์ปดํ“จํ„ฐ๊ฐ€ ์ธ์‹ํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” ๊ธฐ์ˆ ๋กœ, ์ปดํ“จํ„ฐ ๋น„์ „ ๋ถ„์•ผ์—์„œ ํ˜„์žฌ ๋„๋ฆฌ ์“ฐ์ด๋Š” ๋Œ€ํ‘œ์ ์ธ ๊ธฐ์ˆ  ์ค‘ ํ•˜๋‚˜์ด๋‹ค. ๋ณธ ๋Œ€ํšŒ์—์„œ๋Š” OCR task ์ค‘ ๊ธ€์ž ๊ฒ€์ถœ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๊ณ ๋„ํ™”ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ํ•œ๋‹ค.


Dataset

  • ๊ธฐ๋ณธ ๋ฐ์ดํ„ฐ์…‹: ICDAR17_Korean (ICDAR17-MLT ์„œ๋ธŒ์…‹)
  • ์ „์ฒด ์ด๋ฏธ์ง€ ๊ฐœ์ˆ˜
    • Train: 536
    • Test: 300

๋ฉค๋ฒ„

๊น€์˜์šด ์ด์Šนํ˜„ ์ž„์„œํ˜„ ์ „์„ฑํœด ํ—ˆ์„์šฉ
Github Github Github Github Github
  • ๊น€์˜์šด : EDA, Annotation data converting(Polygon to Rectangle)
  • ์ด์Šนํ˜„ : EDA, ์ „์ˆ˜์กฐ์‚ฌ ํ™˜๊ฒฝ ์„ธํŒ…, Fine-tuning ์ ์šฉ
  • ์ž„์„œํ˜„ : EDA, AiHub ์˜คํ”ˆ ๋ฐ์ดํ„ฐ์…‹ ์ ์šฉ
  • ์ „์„ฑํœด : EDA, Augmentation ์‹คํ—˜
  • ํ—ˆ์„์šฉ : EDA, Validate ๊ธฐ๋Šฅ ๊ฐœ๋ฐœ

ํ”„๋กœ์ ํŠธ ์ˆ˜ํ–‰ ๊ฒฐ๊ณผ

2

1. ๋ฐ์ดํ„ฐ ์ „์ˆ˜ ์กฐ์‚ฌ

  • ๋Œ€ํšŒ์—์„œ ์ œ๊ณตํ•œ Train data 536๊ฐœ ์ด๋ฏธ์ง€์™€ annotation tool์„ ํ™œ์šฉํ•˜์—ฌ ์ง์ ‘ labelingํ•œ 1288๊ฐœ์˜ ์ด๋ฏธ์ง€์—์„œ ์ž˜๋ชป ๋ ˆ์ด๋ธ”๋ง๋œ ์ผ€์ด์Šค๊ฐ€ ๋ฐœ๊ฒฌ๋˜์–ด ์ „์ˆ˜ ์กฐ์‚ฌ๋ฅผ ์ง„ํ–‰
  • ์ „์ˆ˜ ์กฐ์‚ฌ ๊ฒฐ๊ณผ ์ž˜๋ชป ๋ ˆ์ด๋ธ”๋ง๋œ ๊ฒฝ์šฐ, ํ•™์Šต์—์„œ ์ œ์™ธํ•˜์˜€๋‹ค.
    • ๋Œ€ํšŒ ๋ฐ์ดํ„ฐ 536๊ฐœ ์ค‘ 16๊ฐœ, ์ถ”๊ฐ€๋กœ ์ œ๊ณต๋ฐ›์€ 1288๊ฐœ ๋ฐ์ดํ„ฐ ์ค‘ 127๊ฐœ์˜ ์ด๋ฏธ์ง€์—์„œ mis-labeling ๋ฐœ๊ฒฌํ•˜์˜€๊ณ , ์ตœ์ข…์ ์œผ๋กœ 143๊ฐœ์˜ ์ด๋ฏธ์ง€๋ฅผ ์ œ๊ฑฐํ•˜์˜€๋‹ค.

2. Text Detection Open Dataset ์กฐ์‚ฌ

  • 300๊ฐœ์˜ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ์— ๋น„ํ•ด ํ•™์Šต ๋ฐ์ดํ„ฐ๊ฐ€ ์ ์–ด ์ถ”๊ฐ€ ๋ฐ์ดํ„ฐ ํ™•๋ณด๊ฐ€ ํ•„์ˆ˜์ ์ด๋ผ๊ณ  ํŒ๋‹จํ•˜์˜€๊ณ , ์ด๋ฅผ ์œ„ํ•ด OCR ๊ด€๋ จ ์˜คํ”ˆ ๋ฐ์ดํ„ฐ์…‹ ์กฐ์‚ฌ ์ง„ํ–‰
  • AiHub์™€ ICDAR19 ๋ฐ์ดํ„ฐ์…‹ ์„ ์ • ํ›„ ๋Œ€ํšŒ ๋ฐ์ดํ„ฐ์…‹๊ณผ ๋น„์Šทํ•˜๊ฒŒ annotation ๋ ์–ด ์žˆ๋Š”์ง€ ๋ถ„์„ํ•˜์˜€๊ณ , ์„ ๋ณ„๋œ ์ด๋ฏธ์ง€๋ฅผ ํ•™์Šต์— ์‚ฌ์šฉํ•˜๊ธฐ๋กœ ๊ฒฐ์ •
  • AiHub ๋ฐ์ดํ„ฐ์…‹์€ ๋ฐ์ดํ„ฐ ์ˆ˜๊ฐ€ ๋งŽ์•˜์ง€๋งŒ ๋Œ€ํšŒ ๋ชฉ์ ์— ๋ถ€ํ•ฉํ•˜๋Š” ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๊ฐ€ ์ ์–ด์„œ ํ•™์Šต ์‹œ ๊ฒฐ๊ณผ๊ฐ€ ์ข‹์ง€ ์•Š์•„ ์ตœ์ข…์ ์œผ๋กœ ์ œ์™ธ

3. Validate

  • Test ๋ฐ์ดํ„ฐ์…‹ ๋ถ„ํฌ์™€ ๋น„์Šทํ•˜๊ฒŒ Validation ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ์„ฑํ•˜์—ฌ ๋ณธ ๋Œ€ํšŒ์˜ ํ‰๊ฐ€์ง€ํ‘œ์ธ DetEval score๊ฐ€ valid-test ๊ฐ„ ์ ์ˆ˜๊ฐ€ alignํ•˜๋„๋ก ์„ธํŒ…
  • ์—ํญ๋งˆ๋‹ค Validation์…‹์˜ Mean loss, Classification loss, Angle loss, IoU loss๋ฅผ ํ™•์ธํ•˜์—ฌ ์ตœ์ ์˜ ๋ชจ๋ธ์„ ์ฐพ์„ ์ˆ˜ ์žˆ๋„๋ก ์‹คํ—˜ ํ™˜๊ฒฝ์„ ๊ตฌ์„ฑ

4. Fine-Tuning ์ „๋žต

  • ์ œ๊ณต๋ฐ›์€ ํ•™์Šต ๋ฐ์ดํ„ฐ๋Š” ํ•œ๊ธ€ ๋ฐ ์˜์–ด ๊ธ€์ž ์˜์—ญ์„ ํฌํ•จํ•˜๋Š” 536๊ฐœ์˜ ์ด๋ฏธ์ง€์ด๋‹ค.
  • ์ด 10,000๊ฐœ์˜ ๋ฐ์ดํ„ฐ(ํ•œ๊ธ€, ์˜์–ด ์™ธ ์—ฌ๋Ÿฌ ์–ธ์–ด ํฌํ•จ)๊ฐ€ ์กด์žฌํ•˜๋Š” ICDAR19๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋จผ์ € ๋ชจ๋ธ์„ ํ•™์Šตํ•œ ํ›„, target dataset์ธ 536๊ฐœ์˜ ํ•™์Šต ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด fine-tuning ์ง„ํ–‰
  • 10,000์—ฌ๊ฐœ์˜ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ๋กœ target dataset์— ๋Œ€ํ•ด fine-tuning์„ ์ทจํ•˜๋Š” ์ „๋žต์€ ๋ชจ๋ธ์˜ ๊ฒฌ๊ณ ์„ฑ์„ ๋”ํ•ด์ฃผ์—ˆ๊ณ , ๋ชจ๋ธ ์„ฑ๋Šฅ์ด ํฌ๊ฒŒ ํ–ฅ์ƒ๋˜์—ˆ๋‹ค.

5. Hyperparameter Tuning

  • Fine-tuning๊ณผ augmentation์œผ๋กœ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ์ตœ์ ํ™”ํ•˜๊ธฐ ์œ„ํ•˜์—ฌ learning rate์„ ๋”์šฑ ๋‚ฎ์ถฐ ํ•™์Šตํ•˜๋Š” ๋“ฑ hyperparameter tuning ์ง„ํ–‰
  • ์ตœ์ข… ๋ฆฌ๋”๋ณด๋“œ 0.6752 ์ ์ˆ˜ ๊ธฐ๋ก(8์œ„/19)

แ„‰แ…ณแ„แ…ณแ„…แ…ตแ†ซแ„‰แ…ฃแ†บ 2022-04-26 แ„‹แ…ฉแ„’แ…ฎ 11 55 00


level2-data-annotation_cv-level2-cv-11's People

Contributors

heoseokyong avatar sseunghyuns avatar

Stargazers

 avatar  avatar  avatar

Forkers

sseunghyuns

level2-data-annotation_cv-level2-cv-11's Issues

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    ๐Ÿ–– Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. ๐Ÿ“Š๐Ÿ“ˆ๐ŸŽ‰

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google โค๏ธ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.