Table of content (short-version) [paper]


Summary

  • ์ผ๋ฐ˜ํ™”๋œ(universal) re-identification ๋…ผ๋ฌธ
    • source๋Š” train/test๋‹ค ์”€(DB์—ฌ๋Ÿฌ๊ฐœ)
    • target๋Š” test๋งŒ (DB์—ฌ๋Ÿฌ๊ฐœ)
    • ํ•™์Šต DB์— ์—†์—ˆ๋˜ ๊ฒƒ์„ inference (universal DA๊ฐ™์€)
    • UDA์™€ ๋‹ค๋ฅธ์ ์€ target domain์œผ๋กœ ์—…๋ฐ์ดํŠธํ•  ํ•„์š” ์—†์Œ
    • Meta-learning์˜ ์ผ์ข…(one-shot/few-shot): memory bank๋ฅผ ์ด์šฉ, target์œผ๋กœ ์—…๋ฐ์ดํŠธ ํ•  ํ•„์š” ์—†์Œ
  • ์ „์ฒด ํ”„๋ ˆ์ž„์›Œํฌ
    • ๊ฐค๋Ÿฌ๋ฆฌ๋Š” ํŠน์ง•์„ ํ•œ๋ฒˆ ๋ฝ‘์€ ๋‹ค์Œ์— ๋‹ค์‹œ ๋งคํ•‘ ์„œ๋ธŒ๋„ท์„ ํ†ตํ•ด classifier weight๋ฅผ ์ƒ์„ฑ
    • ํ”„๋กœ๋ธŒ๋Š” ๊ทธ๋ƒฅ ๋ฐ”๋กœ ํŠน์ง•์„ ์ƒ์„ฑ
    • Encoding subnet, mapping subnet, memory bank๋กœ ๊ตฌ์„ฑ


[Domain-Invariant Mapping Network ์ „์ฒด ํ”„๋ ˆ์ž„์›Œํฌ]

picture

  • Encoding subnet: feature extractor, mobileNet (๋‹ค๋ฅธ๊ฒƒ๋ณด๋‹ค light weight + ์„ฑ๋Šฅ์€ ๋น„์Šท), batch ๊ตฌ์„ฑ์„ ๋‹ค๋ฅด๊ฒŒ (๋ฐ์ดํ„ฐ์…‹ 5๊ฐœ, ID์ˆ˜ 18530, ํ•™์Šต 18๋งŒ) ํ•œ๋ฒˆ์— ํ•™์Šต์‹œํ‚ค๋ ค๋ฉด ์–ด๋ ค์›Œ์„œ, ์ ์€ ID๋กœ ์ƒ˜ํ”Œ๋ง, ID ๋งˆ๋‹ค (ํ•˜๋‚˜๋Š” ๊ฐค๋Ÿฌ๋ฆฌ, ํ•˜๋‚˜๋Š” ํ”„๋กœ๋ธŒ), CE ์ ์šฉ


[Mini-batch sampling ๋ฐฉ์‹]

picture

  • Mapping subnet: classifier weight vector ํ•™์Šต ๋ฐ ์ƒ์„ฑ FC ๋ช‡๊ฐœ ์‚ฌ์šฉํ•ด์„œ ์ƒ์„ฑ. ํ”„๋กœ๋ธŒ ํ•œ๊ฐœ ๋“ค์–ด์˜ค๋ฉด Cb gallery์™€ ๋น„๊ตํ•ด์„œ mapping network ๊ฑฐ์ณ์„œ Cb๊ฐœ์˜ ํ™•๋ฅ ์ด C๊ฐœ๋กœ ํ‚ค์šฐ๊ณ (zero padding) ๊ทผ๋ฐ ์ด๊ฒŒ ๋„ˆ๋ฌด ์ปค์„œ ์ฐจ๋ณ„์„ฑ์ด ๋–จ์–ด์ง„๋‹ค. ๊ทธ๋ž˜์„œ ๋ฉ”๋ชจ๋ฆฌ๋ฑ…ํฌ ๊ฐ€์ ธ์™€์„œ C๊ฐœ๋กœ ์•ˆ๋„˜๊ฒจ๋„ ๋˜๊ฒŒ๋”. ๋ชจ๋“  batch์—์„œ๋Š” CbxCb๋‚˜์˜ด.
  • memory bank: D(feature dim) x C(all class) ๊ณฑํ•˜๋ฉด C๊ฐœ๊ฐ€ ๋œ๋‹ค. ์ฆํญ ์•ˆํ•ด๋„๋œ๋‹ค. ๋ฉ”๋ชจ๋ฆฌ๋ฑ…ํฌ๋Š” ํฐ๋ฐ ์›ํ•˜๋Š” minibatch ๋ถ€๋ถ„๋งŒ ์—…๋ฐ์ดํŠธํ•œ๋‹ค. memory bank๋ฅผ ์‚ฌ์šฉํ•œ ๊ฒฐ๊ณผ์— CE loss์‚ฌ์šฉ. running average. regularization์„ ์ ์šฉํ•ด์„œ ์•ˆ์ •ํ™”. logit-triplet loss.


[DIMN ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜]

picture


[๋ฐ์ดํ„ฐ์…‹ ์ •๋ณด]

picture


[Performance comparison 1]

picture


[Performance comparison 2]

picture


[Ablation study]

picture


[Few-shot learning ๊ฒฐ๊ณผ]

picture


References

[1] Song, Jifei, et al. โ€œGeneralizable Person Re-identification by Domain-Invariant Mapping Network.โ€ Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2019.