๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

๋ถ„๋ฅ˜ ์ „์ฒด๋ณด๊ธฐ31

[Deep learning] Class-Incremental Learning (LwF, PODNet) ๋ชจ๋“  ๋ฐ์ดํ„ฐ๋ฅผ ํ•œ ๋ฒˆ์— ์ €์žฅํ•˜๊ณ  ํ•™์Šตํ•˜๋Š” ๊ฒƒ์€ ๋น„ํšจ์œจ์ ์ด๊ณ  ์–ด์ฉŒ๋ฉด ๋น„ํ˜„์‹ค์ ์ด๋‹ค. ํŠนํžˆ, ๋ฐ์ดํ„ฐ๊ฐ€ ๋งค์šฐ ํฌ๊ฑฐ๋‚˜ ๋ฏผ๊ฐํ•œ ์ •๋ณด๋ฅผ ํฌํ•จํ•˜๋Š” ๊ฒฝ์šฐ์—๋Š” ๋”์šฑ! ๊ทธ๋ž˜์„œ ๊ณ ์•ˆ๋œ Class-Incremental Learning (CIL)์€ ๋ชจ๋ธ์ด ์‹œ๊ฐ„์ด ์ง€๋‚จ์— ๋”ฐ๋ผ ์ ์ง„์ ์œผ๋กœ ์ƒˆ๋กœ์šด ํด๋ž˜์Šค๋ฅผ ํ•™์Šตํ•˜๋Š” ํ•™์Šต๋ฒ•์ด๋‹ค. ์ „ํ†ต์ ์ธ ํ•™์Šต ๋ฐฉ์‹์—์„œ๋Š” ๋ชจ๋“  ํด๋ž˜์Šค๋ฅผ ํ•œ ๋ฒˆ์— ํ•™์Šตํ•˜์ง€๋งŒ, CIL์—์„œ๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ์ ์ง„์ ์œผ๋กœ ์ œ๊ณต๋˜๋ฉฐ ๋ชจ๋ธ์ด ์ƒˆ๋กœ์šด ํด๋ž˜์Šค๋ฅผ ํ•™์Šตํ•  ๋•Œ ์ด์ „์— ํ•™์Šตํ•œ ๋‚ด์šฉ์„ ์žŠ์ง€ ์•Š๋„๋ก ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋‹ค. ์ด๋Š” CIL์ด ๋‹ค๋ฃจ๋Š” Catastrophic Forgetting ๋ฌธ์ œ๋ผ๊ณ ๋„ ๋ถˆ๋ฆฌ๋Š”๋ฐ, ์ด๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ• ์ค‘ LwF์™€ PODNet์„ ๊ฐ„๋‹จํžˆ ์†Œ๊ฐœํ•ด ๋ณด๊ฒ ๋‹ค.๋‘˜์„ ๊ตฌํ˜„ํ•œ ipynb ํŒŒ์ผ์„ ์•„๋ž˜ Github repo.. 2024. 7. 8.
[Deep learning] What is 'Style transfer'? (CVPR 2016) Image Style Transfer Using Convolutional Neural Networks (https://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/Gatys_Image_Style_Transfer_CVPR_2016_paper.pdf)(ECCV 2016)Perceptual Losses for Real-Time Style Transfer and Super-Resolution (https://arxiv.org/pdf/1603.08155.pdf)  โœจ Style Transfer๋ž€? ์ด๋ฏธ์ง€์˜ '์ปจํ…์ธ '๋Š” ๊ทธ๋Œ€๋กœ ๋‘๊ณ  '์Šคํƒ€์ผ'์„ ๋ณ€ํ™˜ํ•˜๋Š” ๊ธฐ์ˆ ์ด๋‹ค. ํŠนํžˆ 2016๋…„์— ๋ฐœํ‘œ๋œ ๋‘ ๋…ผ๋ฌธ, "Image Style Transfe.. 2024. 7. 7.
[Deep learning] Accelerating the Super-Resolution Convolutional Neural Network ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ (ECCV 2016) Accelerating the Super-Resolution Convolutional Neural Network (https://arxiv.org/pdf/1608.00367.pdf)์ด ๋…ผ๋ฌธ์€ ๊ธฐ์กด์˜ Super Resolution CNN(SRCNN)์˜ ์—ฐ์‚ฐ์„ ๊ฐ€์†ํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ์—ฐ๊ตฌํ–ˆ๋‹ค. Super resolution task๋Š” ์ €ํ•ด์ƒ๋„ ์ด๋ฏธ์ง€๋ฅผ ์ž…๋ ฅ๋ฐ›์•„ ๊ณ ํ•ด์ƒ๋„๋กœ ๋ณต์›ํ•˜๋Š” ์ž‘์—…์ด๋‹ค.  1. ๊ธฐ์กด์˜ SRCNN SRCNN์€ Dong et al. (2014)์— ์˜ํ•ด ์ œ์•ˆ๋œ ๋ชจ๋ธ๋กœ,  ๊ธฐ๋ณธ์ ์œผ๋กœ ์„ธ ๊ฐœ์˜ ์ปจ๋ณผ๋ฃจ์…˜ ๋ ˆ์ด์–ด๋กœ ๊ตฌ์„ฑ๋˜์–ด Patch Extraction and Representation, Non-Linear Mapping, Reconstruction์˜ ๊ณผ์ •์„ ํ†ตํ•ด ์ด๋ฏธ์ง€ ํ•ด์ƒ๋„๋ฅผ ๋ณต์›.. 2024. 7. 7.
[CV] ViT, ViViT (Vision Transformer, Video Vision Transformer) https://yoomimi.tistory.com/entry/Attention-Seq2Seq-Transformer [Deep Learning] Attention, Seq2Seq, TransformerVision Transformer๋ฅผ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด ํ•„์ˆ˜์ ์ธ ๊ฐœ๋…๋“ค์„ ํ•œ๋ฐ ์ •๋ฆฌํ•ด๋ณด๋ ค๊ณ  ํ•œ๋‹ค.์šฐ์„  RNN, LSTM, GRU์— ๊ด€ํ•œ ํฌ์ŠคํŒ…์€ ์•„๋ž˜! ์ด ๊ฐœ๋…์„ ์•Œ์•„์•ผ ์ดํ•ดํ•˜๊ธฐ ํŽธํ•˜๋‹ค. https://yoomimi.tistory.com/entry/RNN-LSTM-GRU [Deepyoomimi.tistory.com ์šฐ์„  Attention๊ณผ Transformer์— ๊ด€ํ•œ ์ดํ•ด๊ฐ€ ํ•„์š”ํ•˜๋‹ค.  1. ViT (Vision Transformer)Transformer๊ฐ€ ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ ๋ถ„์•ผ์—์„œ SOTA๋กœ ์“ฐ์ด๋‹ˆ CV ์ชฝ์—์„œ๋„ ์ด๋ฅผ.. 2024. 7. 5.
[CV] Statistical object recognition, PCA/LDA, SVD ๐Ÿ“ Statistical object recognition, PCA/LDA, SVD2024๋…„ ์—ฐ์„ธ๋Œ€ํ•™๊ต ์ปดํ“จํ„ฐ๊ณผํ•™๊ณผ 4ํ•™๋…„ ๊ณผ๋ชฉ์ธ Computer Vision์„ ์ˆ˜๊ฐ•ํ•˜๋ฉฐ...  #1. Object recognition์—์„œ categorization์— ๋Œ€ํ•œ statisticalํ•œ ๊ด€์ ๋ฒ ์ด์ฆˆ ์ •๋ฆฌ (Bayes Rule) ์ด์šฉ: p(zebra | image) = p(image | zebra) p(zebra) ์‚ฌํ›„ ํ™•๋ฅ  (Posterior): p(zebra โˆฃ image)์šฐ๋„ (Likelihood): p(image โˆฃ zebra)์‚ฌ์ „ ํ™•๋ฅ  (Prior): p(zebra) MAP decision (Maximum a Posteriori Decision): ๊ฒฐ๊ตญ ์šฐ๋ฆฌ์˜ ๋ชฉ์ ์€ posterior๊ฐ€ ์ตœ๋Œ€๊ฐ€ ๋˜๋„๋ก ํ•˜๋Š”.. 2024. 6. 9.
[Deep Learning] Attention, Seq2Seq, Transformer Vision Transformer๋ฅผ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด ํ•„์ˆ˜์ ์ธ ๊ฐœ๋…๋“ค์„ ํ•œ๋ฐ ์ •๋ฆฌํ•ด๋ณด๋ ค๊ณ  ํ•œ๋‹ค.์šฐ์„  RNN, LSTM, GRU์— ๊ด€ํ•œ ํฌ์ŠคํŒ…์€ ์•„๋ž˜! ์ด ๊ฐœ๋…์„ ์•Œ์•„์•ผ ์ดํ•ดํ•˜๊ธฐ ํŽธํ•˜๋‹ค. https://yoomimi.tistory.com/entry/RNN-LSTM-GRU [Deep Learning] RNN, LSTM, GRU ์ด์ •๋ฆฌโ˜… (+ํŒ์„œ)RNN(Recurrent Neural Network)์šฐ์„  ๋” ์ต์ˆ™ํ•œ CNN์—์„œ ์ถœ๋ฐœํ•ด๋ณด์ž. CNN์€ input(๋“ค)์„ ์ด์šฉํ•ด output์„ ์˜ˆ์ธกํ•˜๋Š”๋ฐ, ๊ทธ ๊ณผ์ •์—์„œ data๊ฐ€ ์žฌ์‚ฌ์šฉ๋˜์ง€ ์•Š๋Š”๋‹ค. ๋‹น์—ฐํ•˜๋‹ค. CNN์€ input ํ•˜๋‚˜๋ฅผ ํ•œ๊บผ๋ฒˆ์— ๋„ฃ์–ด์ฃผ๊ธฐ ๋•Œ๋ฌธyoomimi.tistory.com   ๋“ค์–ด๊ฐ€๊ธฐ ์ „, ๊ธฐ๊ณ„ ๋ฒˆ์—ญ์˜ ๋ฐœ์ „ ๊ณผ์ •์„ ์•Œ๊ณ  ๊ฐ€๋ฉด ์ข‹๋‹ค.RNN > LSTM > .. 2024. 2. 19.
[Deep Learning] RNN, LSTM, GRU ์ด์ •๋ฆฌโ˜… (+ํŒ์„œ) RNN(Recurrent Neural Network)์šฐ์„  ๋” ์ต์ˆ™ํ•œ CNN์—์„œ ์ถœ๋ฐœํ•ด๋ณด์ž. CNN์€ input(๋“ค)์„ ์ด์šฉํ•ด output์„ ์˜ˆ์ธกํ•˜๋Š”๋ฐ, ๊ทธ ๊ณผ์ •์—์„œ data๊ฐ€ ์žฌ์‚ฌ์šฉ๋˜์ง€ ์•Š๋Š”๋‹ค. ๋‹น์—ฐํ•˜๋‹ค. CNN์€ input ํ•˜๋‚˜๋ฅผ ํ•œ๊บผ๋ฒˆ์— ๋„ฃ์–ด์ฃผ๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค. ํ•˜๋‚˜๋ฅผ ํ•œ๊บผ๋ฒˆ์—? ์ด๋ฏธ์ง€ ํ•˜๋‚˜๋ฅผ ๋„ฃ์„ ๋•Œ ๊ฐ๊ฐ์˜ ํ”ฝ์…€์„ ์ˆœ์„œ๋Œ€๋กœ ๋„ฃ์ง€ ์•Š๊ณ  ํ•œ๋ฒˆ์— Convolution layer๋ฅผ ๋งŒ๋‚˜๊ฒŒ ํ•ด๋ฒ„๋ฆฌ๋Š” ์ผ์„ ์ƒ์ƒํ•ด๋ณด๋ฉด ๋œ๋‹ค. ๋ฌผ๋ก  Convolution layer์˜ kernel size๋•Œ๋ฌธ์— ๋จผ์ € ์ฝํžˆ๋Š” ๋ถ€๋ถ„์ด ์กด์žฌํ•˜์ง€ ์•Š๋Š๋ƒ ์‹ถ์„ ์ˆ˜ ์žˆ์ง€๋งŒ, ๊ทธ ์ˆœ์„œ๊ฐ€ ์ค‘์š”ํ•œ๊ฐ€? ์ ˆ๋Œ€ ๊ทธ๋ ‡์ง€ ์•Š๋‹ค. ์ด๋ฏธ์ง€์—์„œ locality๊ฐ€ ์ค‘์š”ํ•œ ๊ฒƒ์€ sequence๊ฐ€ ์ค‘์š”ํ•œ ๊ฒƒ๊ณผ๋Š” ๋‹ค๋ฅธ ์˜๋ฏธ๋‹ค. RNN์€ sequence data(์‹œ๊ณ„์—ด dat.. 2024. 1. 12.
[HCI] The Alleviation of Perceptual Blindness During Driving in Urban Areas Guided by Saccades Recommendation, IEEE Transactions on Intelligent Transportation Systems (2022) (๋…ผ๋ฌธ ๋ฆฌ๋ทฐ) โœจ ABSTRACT ์šด์ „ ์‹œ ์ธ์ง€์  ๋งน์ ์ด ์ฃผ์š” ๊ตํ†ต์‚ฌ๊ณ  ์›์ธ ์ค‘ ํ•˜๋‚˜๋‹ค. ์ด ๋…ผ๋ฌธ์€ computational visual attention models (CVAMs)์ด ์ธ๊ฐ„์˜ attention mechanism๊ณผ ์œ ์‚ฌํ•˜๊ฒŒ ์ฃผ์˜๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๊ฒƒ์„ ๊ธฐ๋ฐ˜์œผ๋กœ, ๋„์‹œ ๋„๋กœ ํ™˜๊ฒฝ์—์„œ ์šด์ „ ์•ˆ์ „์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•œ saccades strategy recommendation์„ ์‹œ์‚ฌํ•œ๋‹ค. ์ด ๋…ผ๋ฌธ์˜ ์ฐจ๋ณ„์ ์€ ๊ธฐ์กด ์—ฐ๊ตฌ๋“ค์ด driving task์—์„œ visual attention์„ ์ด์šฉํ•œ computational model์„ testํ•  ๋•Œ image๋‚˜ video๋ฅผ ์‚ฌ์šฉํ•œ๋ฐ ๋ฐ˜ํ•ด real-world task๋ฅผ ์ˆ˜ํ–‰์‹œ์ผฐ๋‹ค๋Š” ์ ์ด๋‹ค. โœจ METHOD eye movements๋ฅผ ๋‹ค์Œ์˜ ํŠน์ง•๋“ค์— ๋”ฐ๋ผ ๋ถ„๋ฅ˜ํ•˜๊ณ , ์ด๋ฅผ .. 2024. 1. 12.