Home [Paper] Crowdsourcing Annotations for Visual Object Detection
Post
Cancel

[Paper] Crowdsourcing Annotations for Visual Object Detection

๐Ÿ“Ž Paper: http://vision.stanford.edu/pdf/bbox_submission.pdf


crowdsourcing์„ ์ด์šฉํ•œ bounding-box annotation system์— ๋Œ€ํ•ด ๋‹ค๋ฃจ๋Š” ๋…ผ๋ฌธ์ด๋‹ค. (์Šคํƒ ํฌ๋“œ Fei-Fei ๊ต์ˆ˜๋‹˜ ์—ฐ๊ตฌ์‹ค ๋…ผ๋ฌธ!) ํšŒ์‚ฌ์—์„œ crowdsourcing ์‚ฌ์—…์„ ์ง„ํ–‰ํ–ˆ์–ด์„œ ์ฝ์—ˆ๋˜ ๋…ผ๋ฌธ์ด๋‹ค.


1. Introduction

bounding box annotation์„ crowd-source ํ•˜๊ธฐ ์œ„ํ•œ fully automated, highly accurate, cost-effective ํ•œ ์‹œ์Šคํ…œ

Requirements

  • quality - ๊ฐ bbox๋Š” tight ํ•ด์•ผ ํ•œ๋‹ค.
  • coverage - ๋ชจ๋“  object๋Š” bbox๋ฅผ ๊ฐ€์ ธ์•ผ ํ•œ๋‹ค.

cost-effective(ํ’ˆ์งˆ์€ ๋ณด์žฅํ•˜๋ฉด์„œ ๋น„์šฉ์„ ์ตœ์†Œํ™”) ํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ high quality์™€ complete coverage๋ฅผ ๋‹ฌ์„ฑํ•˜๋Š” ๋ฐฉ๋ฒ•์€?


Three sub-tasks

ํ•œ ์ž‘์—…์ž๊ฐ€ bbox๋ฅผ ๊ทธ๋ฆฌ๊ณ , ๋‹ค๋ฅธ ์ž‘์—…์ž๊ฐ€ bbox์˜ ํ’ˆ์งˆ์„ ํ™•์ธํ•˜๊ณ , ๋˜ ๋‹ค๋ฅธ ์ž‘์—…์ž๊ฐ€ ๋ชจ๋“  object๊ฐ€ bbox๋กœ ํ‘œ์‹œ๋˜์—ˆ๋Š”์ง€ ํ™•์ธํ•˜๋Š” ๋ฐฉ๋ฒ•

  • drawing: ์ด๋ฏธ์ง€์—์„œ ํ•˜๋‚˜์˜ ๋ฌผ์ฒด์— ๋Œ€์‘ํ•˜๋Š” ํ•˜๋‚˜์˜ bbox๋ฅผ ๊ทธ๋ฆฐ๋‹ค.
  • quality verification: bbox๊ฐ€ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ๊ทธ๋ ค์กŒ๋Š”์ง€ ํ™•์ธํ•œ๋‹ค.
  • coverage verification: ๋ชจ๋“  object์— bbox๊ฐ€ ๊ทธ๋ ค์กŒ๋Š”์ง€ ํ™•์ธํ•œ๋‹ค.

    verification ๊ณผ์ •์€ binary answer๊ฐ€ ํ•„์š”ํ•˜๋ฏ€๋กœ ์ด๋ฏธ ์ž˜ ์•Œ๋ ค์ง„ majority voting๊ณผ ๊ฐ™์€ ๋ฐฉ๋ฒ•์„ ์ ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค.



2. Approach

Work Flow

Untitled

  1. drawing task: ํ•˜๋‚˜์˜ object์— ๋Œ€์‘ํ•˜๋Š” ํ•˜๋‚˜์˜ bbox๋ฅผ ๊ทธ๋ฆฐ๋‹ค. (ex. raccoon)
    • worker training (rules)
      1. all visible part & as tightly as possible
      2. include only one
      3. new instance
      4. check the check box when completed
    • qualification test
      • rule์„ ์ œ๋Œ€๋กœ ์ˆ™์ง€ํ–ˆ๋Š”์ง€ ํ…Œ์ŠคํŠธ ์ด๋ฏธ์ง€ ์…‹์„ ํ†ตํ•ด์„œ ํ™•์ธ ํ›„ instant feedback์„ ์ „์†กํ•œ๋‹ค.
      • ์ถฉ๋ถ„ํžˆ tight ํ•˜์ง€ ์•Š์Œ / solicited object๊ฐ€ ์•„๋‹˜ / ์ด๋ฏธ bbox๊ฐ€ ์žˆ๋Š” object ์ž„
    • ๊ทธ ํ›„ ์‹ค์ œ ์ด๋ฏธ์ง€์—์„œ ์ž‘์—…ํ•  ์ˆ˜ ์žˆ๋‹ค.
  2. quality verification task: ์ƒˆ๋กญ๊ฒŒ ๊ทธ๋ ค์ง„ bbox์˜ quality๋ฅผ ์ธก์ •ํ•˜๊ณ  good bbox๋Š” DB์—, bad bbox๋Š” ๋ฒ„๋ฆฐ๋‹ค.
    • ์ž‘์—…์ž๊ฐ€ ์ง‘์ค‘ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ด๋ฏธ์ง€ ๋‹น ํ•˜๋‚˜์˜ bbox๋งŒ์„ ๋ณด์—ฌ์ค€๋‹ค.
    • worker training (rules)
      1. include an instance of the required object
      2. all visible part & as tightly as possible
      3. include only one
    • qualification test
    • quality control (gold standard)
      • good bbox๋ฅผ bad๋กœ, bad bbox๋ฅผ good๋กœ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒฝ์šฐ๋ฅผ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•จ์ด๋‹ค.
      • batch์— ์ผ๋ถ€ ํฌํ•จ๋˜๋Š” validation images๋ฅผ ์ž‘์—…์ž๊ฐ€ ์˜ฌ๋ฐ”๋ฅด๊ฒŒ ํ‰๊ฐ€ํ•ด์•ผ ์ž‘์—… ๋‚ด์—ญ์ด accept ๋œ๋‹ค.

        validation image๋ฅผ ์œ„ํ•œ good & bad bbox๋ฅผ ์–ป๋Š” ๋ฐฉ๋ฒ•

        • bad bbox๋Š” good bbox๋ฅผ ๋ณ€ํ˜•ํ•ด์„œ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค.
        • good bbox๋Š” majority voting์„ ํ†ตํ•ด ๋ชจ์„ ์ˆ˜ ์žˆ๋‹ค.
          1. ํŠน์ • object๋ฅผ ํฌํ•จํ•˜๋Š” image์—์„œ ์ผ๋ถ€๋ฅผ ์ƒ˜ํ”Œ๋งํ•˜๊ณ  bbox๋ฅผ ์–ป๋Š”๋‹ค.
          2. ์—ฌ๋Ÿฌ ์ž‘์—…์ž๋“ค์ด bbox์— ํ‰์ ์„ ๋งค๊ธฐ๊ณ , strong consensus(at least 3 workers)๊ฐ€ ์žˆ๋Š” ๊ฒƒ๋“ค์„ โ€œgold standardโ€๋กœ ์„ ์ •ํ•œ๋‹ค.
  3. coverage verification task: raccoon์— ํ•ด๋‹นํ•˜์ง€๋งŒ ์•„์ง bbox๋กœ ํ‘œ์‹œ๊ฐ€ ๋˜์ง€ ์•Š์€ object๊ฐ€ ์žˆ๋Š”์ง€ ํ™•์ธํ•˜๊ณ , ๋ชจ๋‘ ํ‘œ์‹œ๊ฐ€ ๋˜์—ˆ์œผ๋ฉด ์™„๋ฃŒํ•œ๋‹ค.
    • ๋ชจ๋“  instance๊ฐ€ bbox๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ๋Š”์ง€ ํ™•์ธํ•œ๋‹ค.
    • ๊ฐ™์€ object๋ฅผ ํฌํ•จํ•˜๋Š” ์ด๋ฏธ์ง€๋“ค์ด ํ•œ ๋ช…์˜ annotator์—๊ฒŒ ๋ฐฐ์ •๋œ๋‹ค.
    • ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ worker training, qualification test๋ฅผ ์ˆ˜ํ–‰ํ•œ๋‹ค.
    • quality control
      • ๋‘ ์ข…๋ฅ˜์˜ validation images๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค.
        • completely covered๋Š” majority voting์„ ํ†ตํ•ด ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค.
        • ๊ทธ๋ ‡์ง€ ์•Š์€ ๊ฒƒ์€ bbox์˜ subset์„ ์ง€์šฐ๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค.
  4. ๋ชจ๋“  raccoon์ด bbox๋กœ ํ‘œ์‹œ๋  ๋•Œ๊นŒ์ง€ ๋ฐ˜๋ณตํ•œ๋‹ค.


Two principles

  • simple (draw only one bbox)
  • have a fixed and predictable amount of work



3. Experiments

20,000๊ฐœ์˜ ์นดํ…Œ๊ณ ๋ฆฌ๋ฅผ ๊ฐ€์ง„ ImageNet ๋ฐ์ดํ„ฐ์…‹์—์„œ 10๊ฐœ์˜ ์นดํ…Œ๊ณ ๋ฆฌ๋ฅผ ์„ ์ •ํ•˜์˜€๊ณ , ๊ฐ ์นดํ…Œ๊ณ ๋ฆฌ ๋‹น 200๊ฐœ์˜ ์ด๋ฏธ์ง€๋ฅผ randomly sample ํ•˜์˜€๋‹ค.

Overall quality

  • image level: 97.9% images๊ฐ€ completed covered
  • bbox level: 99.2% bboxes๊ฐ€ accurate
  • ํ•ด๋‹น ์‹œ์Šคํ…œ์„ ํ†ตํ•ด highly accurate bbox๊ฐ€ ์ƒ์„ฑ๋˜์—ˆ๋‹ค.


Overall cost

cost = ์ž‘์—…์ž๊ฐ€ ์†Œ๋น„ํ•œ ์‹œ๊ฐ„

  • drawing task๊ฐ€ quality/coverage verification task ๋ณด๋‹ค 2๋ฐฐ ์ด์ƒ ์˜ค๋ž˜ ๊ฑธ๋ฆฐ๋‹ค.

    verification task์˜ ๊ฒฝ์šฐ, binary answer๋งŒ ํ•„์š”๋กœ ํ•˜๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค.

  • ํ•ด๋‹น ์‹œ์Šคํ…œ๊ณผ consensus based ๋ฐฉ๋ฒ•์˜ cost ๋น„๊ต

    ย our approachconsensus based approachhow expensive
    mean88.0 sec116.9 sec32.8%
    median42.4 sec58.8 sec38.9%

Untitled


Quality control

  • drawing task: quality verification task๋ฅผ ํ†ตํ•ด control
    • acceptance ratio = 62.2%
  • quality verification task: majority voting ๋ฐฉ๋ฒ•์„ ์ด์šฉํ•œ โ€œgold standardโ€๋ฅผ ํ†ตํ•ด control
    • validation images๋ฅผ ๋Œ€์ƒ์œผ๋กœ performance๋ฅผ ํ‰๊ฐ€ํ•œ๋‹ค.
    • acceptance ratio = 89.9%
  • coverage verification task: quality verification task์™€ ๋น„์Šท
    • validation images๋ฅผ ๋Œ€์ƒ์œผ๋กœ performance๋ฅผ ํ‰๊ฐ€ํ•œ๋‹ค.
    • accpetance ratio = 95.0%

drawing task๋Š” ๋” time consuming ํ•  ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ๋” difficult ํ•˜๋‹ค.


Effectiveness of worker training

drawing task์—์„œ worker training ๊ณผ์ •์„ ์‚ญ์ œํ•˜์˜€์„ ๋•Œ์— ๋น„ํ•ด worker training ๊ณผ์ •์„ ์ง„ํ–‰ํ•  ๋•Œ์˜ quality verification acceptance๊ฐ€ 4.2% ๋†’์•˜๋‹ค.

Untitled

This post is licensed under CC BY 4.0 by the author.

[Tips] ์‚ฌ์šฉ ์ค‘์ธ ํฌํŠธ ๋ฒˆํ˜ธ ํ™•์ธํ•˜๊ธฐ (for Mac)

[Docker] Container ๋‚ด๋ถ€์—์„œ ์‚ฌ์šฉ ์ค‘์ธ ํฌํŠธ ๋ฒˆํ˜ธ ํ™•์ธํ•˜๊ธฐ