๐ŸŒŒ Deep Learning/๋…ผ๋ฌธ ๋ฆฌ๋ทฐ [KOR]

[๋”ฅ๋Ÿฌ๋‹ ๋…ผ๋ฌธ๋ฆฌ๋ทฐ] Class-Balanced Loss Based on Effective Number of Samples (CVPR 2019)

๋ณต๋งŒ 2020. 12. 17. 13:44

 

CVPR 2019์— ๋ฐœํ‘œ๋œ ๋…ผ๋ฌธ์ธ

Class-Balanced Loss Based on Effective Number of Samples

๋ฅผ ์ •๋ฆฌํ•œ ๊ธ€์ž…๋‹ˆ๋‹ค.

 

๋ฐ์ดํ„ฐ์…‹์˜ Class Imbalance๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด ์ƒˆ๋กœ์šด Loss Design๋ฅผ ์ œ์•ˆํ•˜๋Š” ๋…ผ๋ฌธ์ž…๋‹ˆ๋‹ค.

 

 


Long Tailed Dataset

 

Long Tailed Dataset

 

์œ„ ๊ทธ๋ฆผ์€ ๊ฐ Class์— ์†ํ•˜๋Š” Sample์˜ ๊ฐฏ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋‚ธ ๊ทธ๋ž˜ํ”„์ธ๋ฐ์š”,

 

์ผ๋ถ€ ๋ช‡ ๊ฐœ์˜ Class์—๋งŒ Sample๋“ค์ด ๋ชฐ๋ ค ์žˆ๊ณ ,

๋Œ€๋ถ€๋ถ„์˜ Class์—๋Š” ๋งค์šฐ ์ ์€ ์ˆ˜์˜ Sample์ด ์žˆ๋Š” ๋ฐ์ดํ„ฐ์…‹์„ Long Tailed Dataset์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค.

 

์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ์…‹์„ ๊ฐ€์ง€๊ณ  ํ•™์Šตํ•œ ๋ชจ๋ธ์€ ์ผ๋ฐ˜์ ์œผ๋กœ ์„ฑ๋Šฅ์ด ์ž˜ ๋‚˜์˜ค์ง€ ์•Š๋Š”๋ฐ,

Large-scale, real-world ๋ฐ์ดํ„ฐ์…‹๋“ค์€ ๋ณดํ†ต Long Tailed ํ˜•ํƒœ๋ฅผ ๋„๊ณ  ์žˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค.

 

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์ด๋Ÿฌํ•œ Long Tailed Dataset์„ ํšจ๊ณผ์ ์œผ๋กœ ํ•™์Šต์‹œํ‚ค๊ธฐ ์œ„ํ•œ Loss design์„ ์ œ์•ˆํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

 

 

 


Existing Strategies

Long Tailed Dataset์„ ํšจ๊ณผ์ ์œผ๋กœ ํ•™์Šต์‹œํ‚ค๊ธฐ ์œ„ํ•œ ์—ฐ๊ตฌ๋Š” ์ด์ „์—๋„ ๋งŽ์ด ์ง„ํ–‰๋˜์–ด ์™”์Šต๋‹ˆ๋‹ค.

 

๊ธฐ์กด์˜ ๋ฐฉ๋ฒ•๋“ค์€ ํฌ๊ฒŒ ๋‘ ๊ฐ€์ง€ ์นดํ…Œ๊ณ ๋ฆฌ๋กœ ๋‚˜๋ˆŒ ์ˆ˜ ์žˆ๋Š”๋ฐ์š”,

Re-sampling ๊ณผ Re-weighting ์ž…๋‹ˆ๋‹ค.

 

 

Re-sampling์€ ๋ฐ์ดํ„ฐ์…‹์„ ์ˆ˜์ •ํ•ด Class Imbalance๋ฅผ ํ•ด์†Œํ•˜๋Š” ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค.

Minor Class์˜ ์ƒ˜ํ”Œ์„ ์ค‘๋ณตํ•˜์—ฌ ์‚ฌ์šฉํ•˜๋Š” Over-sampling์ด ์žˆ๊ณ ,

Major Class์˜ ์ƒ˜ํ”Œ์„ ์ผ๋ถ€ ์ œ๊ฑฐํ•˜์—ฌ ์‚ฌ์šฉํ•˜๋Š” Under-sampling์ด ์žˆ์Šต๋‹ˆ๋‹ค.

 

ํ•˜์ง€๋งŒ ๋‹น์—ฐํžˆ ์ด๋Ÿฌํ•œ ๋ฐฉ๋ฒ•๋“ค์€ ๊ทผ๋ณธ์ ์ธ ํ•ด๊ฒฐ์ฑ…์ด ๋˜์ง€๋Š” ๋ชปํ•ฉ๋‹ˆ๋‹ค.

Over-sampling์€ overfitting์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๊ณ , under-sampling์€ ๊ท€์ค‘ํ•œ sample๋“ค์„ ๋ฒ„๋ฆด ์ˆ˜ ์žˆ๋‹ค๋Š” ๋‹จ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค.

 

 

๋”ฐ๋ผ์„œ Re-weighting์˜ ๋ฐฉํ–ฅ์œผ๋กœ ๋งŽ์€ ์—ฐ๊ตฌ๊ฐ€ ์ง„ํ–‰๋˜์—ˆ๋Š”๋ฐ์š”,

์ด๋Š” Minor Class์˜ Loss์— ๋” ํฐ ๊ฐ€์ค‘์น˜๋ฅผ ์ฃผ๋Š” ๋ฐฉ์‹์ž…๋‹ˆ๋‹ค.

(1) Class frequency์— ๋ฐ˜๋น„๋ก€ํ•˜๋Š” ๊ฐ€์ค‘์น˜๋ฅผ ์ถ”๊ฐ€ํ•ด์ฃผ๋Š” ๋ฐฉ์‹๊ณผ,

์ด๋ฅผ ๋ฐœ์ „์‹œ์ผœ (2) Class frequency์˜ Square root์— ๋ฐ˜๋น„๋ก€ํ•˜๋Š” ๊ฐ€์ค‘์น˜๋ฅผ ์ถ”๊ฐ€ํ•ด์ฃผ๋Š” ๋ฐฉ์‹์ด ์ œ์•ˆ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

 

 

 

 

๋ณธ ๋…ผ๋ฌธ์—์„œ ์ œ์•ˆํ•˜๋Š” ๋ฐฉ์‹ ์—ญ์‹œ Re-weighting์˜ ํ•œ ์ข…๋ฅ˜์ž…๋‹ˆ๋‹ค.

๋‹ค๋งŒ, ๊ธฐ์กด ๋ฐฉ๋ฒ•๋“ค๊ณผ ๋‹ค๋ฅด๊ฒŒ Sample์˜ Effective number์— ๋ฐ˜๋น„๋ก€ํ•˜๋Š” ๊ฐ€์ค‘์น˜๋ฅผ ์ถ”๊ฐ€ํ•ด์ฃผ๋Š” ๋ฐฉ์‹์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค.

 

 

 


Effective Number of Samples

๋ณธ ๋…ผ๋ฌธ์€ Effective Number of Samples๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ Loss์— ๊ฐ€์ค‘์น˜๋ฅผ ๋ถ€์—ฌํ•ฉ๋‹ˆ๋‹ค.

๊ทธ๋ ‡๋‹ค๋ฉด Effective Number์ด๋ž€ ๋ฌด์—‡์ผ๊นŒ์š”?

 

๊ฐ„๋‹จํ•˜๊ฒŒ ๋งํ•˜๋ฉด Data ๊ฐ„์˜ Information Overlap์„ ๋‚˜ํƒ€๋‚ด๋Š” ์ˆ˜์น˜์ž…๋‹ˆ๋‹ค.

 

Sample์˜ ์ˆ˜๊ฐ€ ๋Š˜์–ด๋‚ ์ˆ˜๋ก, Model์ด Data๋กœ๋ถ€ํ„ฐ ์–ป์„ ์ˆ˜ ์žˆ๋Š” Marginal benefit์€ ์ค„์–ด๋“ญ๋‹ˆ๋‹ค.

์ฒ˜์Œ ๋ช‡ ๊ฐœ์˜ Sample๋งŒ ์žˆ์„ ๋•Œ์—๋Š” ๊ฐ๊ฐ์˜ Sample๋“ค์ด ์„œ๋กœ ๋‹ค๋ฅธ ์ •๋ณด๋ฅผ ์•Œ๋ ค์ฃผ์ง€๋งŒ,
Sample์˜ ์ˆ˜๊ฐ€ ๋งŽ์•„์งˆ์ˆ˜๋ก ๊ฐ Sample์ด ๊ฐ€์ง€๊ณ  ์žˆ๋Š” ์ •๋ณด ๊ฐ„์˜ Overlap์ด ๋ฐœ์ƒํ•˜๊ธฐ ๋•Œ๋ฌธ์—,
Model์ด ๋ฐฐ์šฐ๋Š” ์ƒˆ๋กœ์šด ์ •๋ณด์˜ ์–‘์€ ์ ์  ์ค„์–ด๋“ ๋‹ค๋Š” ๊ฒƒ์ด์ฃ .

 


 

 

Effective Number์„ ๊ตฌํ•˜๊ธฐ ์œ„ํ•ด Random Covering Problem์„ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค.

 

 

  • ์–ด๋–ค Class์— ํ•ด๋‹นํ•˜๋Š” ๋ชจ๋“  Data๋“ค์˜ ์ง‘ํ•ฉ์„ set S๋ผ๊ณ  ํ•˜๊ณ , ์ด set์˜ Volume์„ N์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค.
  • ๊ฐ๊ฐ์˜ Sample์€ set S์˜ subset์— ํ•ด๋‹นํ•˜๊ณ , Volume์€ 1์ž…๋‹ˆ๋‹ค.
  • Sample๋“ค์€ ์„œ๋กœ Overlapํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
  • Sample๋“ค์€ Randomํ•˜๊ฒŒ ์ถ”์ถœ๋˜๊ณ , ์ „์ฒด set S๋ฅผ Coverํ•˜๋Š” ๊ฒƒ์ด ๋ชฉํ‘œ์ž…๋‹ˆ๋‹ค.

 

Sample๋“ค์ด Randomํ•˜๊ฒŒ ์ถ”์ถœ๋  ๋•Œ, ์ด๋“ค์ด Coverํ•˜๋Š” Volume์€ ์ ์  ์ฆ๊ฐ€ํ•ฉ๋‹ˆ๋‹ค.

 

ํ•˜์ง€๋งŒ Sample๋“ค ๊ฐ„์˜ Overlap์ด ๋ฐœ์ƒํ•˜๊ณ , Sample ์ˆ˜๊ฐ€ ๋ฐœ์ƒํ• ์ˆ˜๋ก Overlapํ•  ํ™•๋ฅ ์€ ์ฆ๊ฐ€ํ•˜๊ธฐ ๋•Œ๋ฌธ์—

Sample์˜ ์ˆ˜๊ฐ€ ๋Š˜์–ด๋‚ ์ˆ˜๋ก Volume์˜ ๊ธฐ๋Œ€๊ฐ’์€ ์ ์  ์ค„์–ด๋“ญ๋‹ˆ๋‹ค.

 

 

์—ฌ๊ธฐ์„œ Effective Number์„ Sample๋“ค์˜ Expected Volume์œผ๋กœ ์ •์˜ํ•ฉ๋‹ˆ๋‹ค.

 


 

Effective Number์˜ ๊ณ„์‚ฐ์„ ์œ„ํ•ด ๋ฌธ์ œ๋ฅผ ์ข€ ๋” ๋‹จ์ˆœํ™”์‹œํ‚ค๊ฒ ์Šต๋‹ˆ๋‹ค.

 

  • Partial Overlapping์€ ์กด์žฌํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
  • ์ฆ‰, ์„œ๋กœ ๋‹ค๋ฅธ ๋‘ ๊ฐœ์˜ Sample์€ ์™„์ „ํžˆ Overlappingํ•˜๊ฑฐ๋‚˜, ์„œ๋กœ Overlappingํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

 

 

 

 

์œ„์™€ ๊ฐ™์€ ์กฐ๊ฑด์„ ์ถ”๊ฐ€ํ•˜๋ฉด Overlapping์ด ๋ฐœ์ƒํ•  ํ™•๋ฅ ๊ณผ, ๋ฐœ์ƒํ•˜์ง€ ์•Š์„ ํ™•๋ฅ ์„ ๊ตฌํ•  ์ˆ˜ ์žˆ์–ด

n๊ฐœ์˜ Sample์„ ์ถ”์ถœํ–ˆ์„ ๋•Œ์˜ Effective Number์„ ์•„๋ž˜์™€ ๊ฐ™์ด ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

 

 

n๊ฐœ์˜ Sample์„ ์ถ”์ถœํ–ˆ์„ ๋•Œ์˜ Effective Number

 

 

์ฆ๋ช… ๊ณผ์ •์€ ๊ฐ„๋‹จํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์›๋ฌธ์„ ์ฒจ๋ถ€ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

 

 

 

 

 

Effective Number์— ๋Œ€ํ•œ ์‹์„ ํ’€์–ด ์“ฐ๋ฉด ์•„๋ž˜์™€ ๊ฐ™์ด ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

์ด๋Š” i๋ฒˆ์งธ ์ถ”์ถœํ•œ Sample์ด Effective Number์— ๐›ฝ^i ๋งŒํผ ๊ธฐ์—ฌํ•จ์„ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.

 

 

 

 

 


Class-Balanced Loss

๋“œ๋””์–ด ์ด ๋…ผ๋ฌธ์—์„œ ์ œ์•ˆํ•˜๋Š” Loss Design๊นŒ์ง€ ์™”์Šต๋‹ˆ๋‹ค.

 

๋ฐ”๋กœ.. Loss์— ๊ฐ Class์˜ Effective Number์— ๋ฐ˜๋น„๋ก€ํ•˜๋Š” ๊ฐ€์ค‘์น˜๋ฅผ ๊ณฑํ•ด ์ฃผ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.

 


 

๋‹ค์Œ์€ i๋ฒˆ์งธ Class์˜ Effective Number์ธ๋ฐ์š”, (์œ„์—์„œ ๊ตฌํ•œ ๊ฐ’๊ณผ ๋™์ผํ•ฉ๋‹ˆ๋‹ค)

์ด ๋•Œ n_i๋Š” Class i์— ์†ํ•˜๋Š” Sample์˜ ๊ฐœ์ˆ˜๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค.

 

 

i๋ฒˆ์งธ Class์˜ Effective Number

 

 

 

์‹ค์ œ๋กœ๋Š” N์˜ ๊ฐ’์„ ์•Œ ์ˆ˜ ์—†๊ธฐ ๋•Œ๋ฌธ์— ์ ์ ˆํ•œ ๊ฐ’์„ ์ฐพ์•„์•ผ ํ•˜๋Š”๋ฐ, 

๊ฐ๊ฐ์˜ Class๋งˆ๋‹ค N๊ฐ’์„ ๋‹ค๋ฅด๊ฒŒ ํ•˜๋ฉด Parameter์˜ ๊ฐœ์ˆ˜๊ฐ€ ๋„ˆ๋ฌด ๋งŽ์•„์ง‘๋‹ˆ๋‹ค.

 

 

๋”ฐ๋ผ์„œ ๋ชจ๋“  Class์˜ N_i ๊ฐ’์„ ํ†ต์ผ์‹œ์ผœ ์ฃผ๋ฉด i๋ฒˆ์งธ Class์˜ Effective Number์€ ๋‹ค์Œ๊ณผ ๊ฐ™๊ฒŒ ๋ฉ๋‹ˆ๋‹ค.

 

 

N์„ ํ†ต์ผ์‹œํ‚จ ํ›„ i๋ฒˆ์งธ Class์˜ Effective Number

 

 


 

๊ทธ๋ฆฌ๊ณ  ์ด ๊ฐ’์„ ์ด์šฉํ•œ Class-Balanced(CB) Loss๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.

 

 

Class-Balanced Loss

 

 

 

์ ์ ˆํ•œ ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ ๐›ฝ ๊ฐ’์„ ์„ ํƒํ•˜์—ฌ ์ฃผ๋ฉด ๋˜๊ณ ,

Model๊ณผ Loss์˜ ์ข…๋ฅ˜์— ๋ฌด๊ด€ํ•˜๋‹ค๋Š” ์žฅ์ ์ด ์žˆ์Šต๋‹ˆ๋‹ค.

 

 


Experiments

์‹คํ—˜์— ์‚ฌ์šฉํ•œ Loss์˜ ์ข…๋ฅ˜๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.

  • Softmax Cross-entropy Loss
  • Sigmoid Cross-entropy Loss
  • Focal Loss

 

์ด ์„ธ๊ฐ€์ง€ ๋ฐ์ดํ„ฐ์…‹์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.

  • CIFAR-10๊ณผ CIFAR-100์„ ์ด์šฉํ•ด ์ง์ ‘ ๋งŒ๋“  Long Tailed Dataset (Imbalance์˜ ์ •๋„๋ฅผ 5๊ฐ€์ง€๋กœ ํ•˜์—ฌ ๋งŒ๋“ฆ)
  • iNaturalist 2017 & 2018 (Real-world Long Tailed Dataset ์ด๋ผ๊ณ  ํ•˜๋„ค์š”)
  • ImageNet (Long Tailed๊ฐ€ ์•„๋‹Œ Dataset์—๋„ ์‹คํ—˜์„ ์ง„ํ–‰ํ•ด ๋ณด์•˜์Šต๋‹ˆ๋‹ค)

 

CIFAR-100์„ ์ด์šฉํ•ด ๋งŒ๋“  Long Tailed Dataset

 

 


 

Long Tailed CIFAR-10๊ณผ CIFAR-100์— ๋Œ€ํ•œ ์‹คํ—˜ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค.

 

๐›ฝ๊ฐ’์œผ๋กœ๋Š” 0.9, 0.99, 0.999, 0.9999 ๋„ค ๊ฐ€์ง€๋ฅผ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.

 

๊ฐ Loss์— ๋Œ€ํ•œ Error rate๋ฅผ ๋‚˜ํƒ€๋ƒˆ์œผ๋ฉฐ

์œ„๋Š” ์ผ๋ฐ˜์ ์ธ Loss๋ฅผ ์ด์šฉํ–ˆ์„ ๋•Œ,

์•„๋ž˜๋Š” Class-Balanced Loss๋ฅผ ์ด์šฉํ–ˆ์„ ๋•Œ์˜ ๊ฒฐ๊ณผ์ด๋ฉฐ ๊ฐ€์žฅ ๊ฒฐ๊ณผ๊ฐ€ ์ข‹์•˜๋˜ Loss์™€ Hyperparameter๋งŒ์„ ๊ธฐ์žฌํ–ˆ์Šต๋‹ˆ๋‹ค.

 

 

 

 

๋ชจ๋“  ๊ฒฝ์šฐ์—์„œ CB Loss๊ฐ€ ๋” ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์คฌ์Šต๋‹ˆ๋‹ค.

 

ํŠน์ดํ•œ ์ ์œผ๋กœ๋Š”, ๋ณดํ†ต ๋น„์ „ ์˜์—ญ์—์„œ Softmax Loss๊ฐ€ ์ฃผ๋กœ ์‚ฌ์šฉ๋˜๋Š”๋ฐ, CB Loss๋ฅผ ์ด์šฉํ•œ ๊ฒฐ๊ณผ์—์„œ๋Š” Sigmoid์™€ Focal Loss๊ฐ€ ์„ฑ๋Šฅ์ด ๋” ์ข‹์•˜๊ณ ,

CIFAR-10 ๋ฐ์ดํ„ฐ์…‹์—์„œ๋Š” ๐›ฝ=0.9999์ผ ๋•Œ ๊ฐ€์žฅ ์„ฑ๋Šฅ์ด ์ข‹์•˜์œผ๋‚˜, CIFAR-100 ๋ฐ์ดํ„ฐ์…‹์—์„œ๋Š” Imbalance์˜ ์ •๋„์— ๋”ฐ๋ผ ๊ฐ๊ธฐ ๋‹ค๋ฅธ ๐›ฝ๊ฐ’์—์„œ ์ตœ๊ณ  ์„ฑ๋Šฅ์ด ๋‚˜์™”๋‹ค๊ณ  ํ•ด์š”.

 

 


 

 

 

 

 

๐›ฝ๊ฐ’์— ๋”ฐ๋ฅธ ์„ฑ๋Šฅ์„ ํ‘œ๋กœ ๋‚˜ํƒ€๋‚ธ ๊ฒƒ์ž…๋‹ˆ๋‹ค.

 

CIFAR-10์˜ ๊ฒฝ์šฐ ๐›ฝ๊ฐ’์ด ํด์ˆ˜๋ก ์„ฑ๋Šฅ์ด ๋” ์ข‹์•„์กŒ๊ณ ,

CIFAR-100์˜ ๊ฒฝ์šฐ ์ž‘์€ ๐›ฝ๊ฐ’์„ ์‚ฌ์šฉํ–ˆ์„ ๋•Œ๋งŒ ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ์žˆ์—ˆ๋‹ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค.

 


 

๊ทธ ์ด์œ ๋ฅผ ์•Œ์•„๋ด…์‹œ๋‹ค..

 

 

๋ถ„์„์ด ์ •๋ง ๋๋„ ์—†๋‹ค.. ์—ด์‹ฌํžˆ ๋…ผ๋ฌธ์„ ์“ฐ์…จ๊ตฐ์š”

 

 

 

Effective Number์˜ ๊ณ„์‚ฐ์‹์„ ๋‹ค์‹œ ๊ฐ€์ ธ์™€ ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.

 

Effective Number์˜ ๊ณ„์‚ฐ์‹

 

 

N์ด ์˜๋ฏธํ•˜๋Š” ๊ฒƒ์€ unique prototype์˜ ์ˆ˜๋ผ๊ณ  ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

๋ณด๋‹ค Fineํ•œ ๋ฐ์ดํ„ฐ์…‹์€ Coarseํ•œ ๋ฐ์ดํ„ฐ์…‹๋ณด๋‹ค ์ž‘์€ N ๊ฐ’์„ ๊ฐ€์ง„๋‹ค๊ณ  ์ƒ๊ฐํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

 

CIFAR-100์˜ Class๋“ค์€ CIFAR-10๋ณด๋‹ค ๋” Fineํ•˜๊ธฐ ๋•Œ๋ฌธ์— ๋” ์ž‘์€ N ๊ฐ’์„ ๊ฐ€์งˆ ๊ฒƒ์ด๊ณ , (์˜ˆ๋ฅผ ๋“ค์ž๋ฉด CIFAR-100์€ ์ฐธ์ƒˆ, ๋น„๋‘˜๊ธฐ,.. ์ด๋Ÿฐ Class๋“ค์ด ์žˆ๋‹ค๋ฉด CIFAR-10์€ ์ƒˆ ๋ผ๋Š” ๋” ํฐ ๋ฒ”์œ„์˜ Class๋“ค๋กœ ์ด๋ฃจ์–ด์ ธ ์žˆ๊ธฐ ๋•Œ๋ฌธ์— CIFAR-100์˜ Class๊ฐ€ ๋” Fineํ•˜๋‹ค๋Š” ๊ฒƒ)๋”ฐ๋ผ์„œ ๋” ์ž‘์€ ๐›ฝ๊ฐ’์—์„œ ์ž˜ ๋™์ž‘ํ•œ๋‹ค.. ๋ผ๊ณ  ํ•ด์„ํ•  ์ˆ˜ ์žˆ๊ฒ ์Šต๋‹ˆ๋‹ค.

 


 

iNaturalist์™€ ILSVRC ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•œ ์‹คํ—˜๊ฒฐ๊ณผ๋ฅผ ๋งˆ์ง€๋ง‰์œผ๋กœ ๋งˆ๋ฌด๋ฆฌํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.

์—ญ์‹œ CB Loss์—์„œ ์ข‹์€ ๊ฒฐ๊ณผ๋ฅผ ๋ƒˆ์Šต๋‹ˆ๋‹ค.

 

 

 

 

 

 


 

Loss์— ๊ฐ€์ค‘์น˜๋ฅผ ์ถ”๊ฐ€ํ•ด ์ฃผ๋Š” ๋ฐฉ์‹์œผ๋กœ ๊ฐ„๋‹จํ•˜๊ฒŒ ๊ตฌํ˜„์ด ๊ฐ€๋Šฅํ•˜๊ธฐ ๋•Œ๋ฌธ์—
์‹ค์ œ๋กœ long tailed dataset์„ ํ•™์Šตํ•ด์•ผ ํ•  ์ผ์ด ์žˆ์„ ๊ฒฝ์šฐ ํ•œ ๋ฒˆ ์‹œ๋„ํ•ด ๋ณด์•„๋„ ์ข‹์„ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹คโ—๏ธ

๋ฐ˜์‘ํ˜•