๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

์ „์ฒด ๊ธ€122

[DL/WIL] 3 ํ…์ŠคํŠธ ๋ฐ์ดํ„ฐ๋ฅผ ์œ„ํ•œ ์ธ๊ณต ์‹ ๊ฒฝ๋ง ์ž…๋ ฅ ๋ฐ์ดํ„ฐ์˜ ํ๋ฆ„์ด ์•ž์œผ๋กœ๋งŒ ์ „๋‹ฌ๋˜๋Š” ์‹ ๊ฒฝ๋ง์„ ํ”ผ๋“œํฌ์›Œ๋“œ ์‹ ๊ฒฝ๋ง(feedforward neural network)๋ผ๊ณ  ํ•œ๋‹ค.Fully connected neural network ๋‚˜ Convolution neural network ๋Š” ๊ธฐ์–ต ์žฅ์น˜๊ฐ€ ์—†์–ด์„œ, ํ•˜๋‚˜์˜ ์ƒ˜ํ”Œ(๋˜๋Š” ํ•˜๋‚˜์˜ ๋ฐฐ์น˜)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ •๋ฐฉํ–ฅ ๊ณ„์‚ฐ์„ ์ˆ˜ํ–‰ํ•˜๊ณ  ๋‚˜๋ฉด ๊ทธ ์ƒ˜ํ”Œ์€ ๋ฒ„๋ ค์ง€๊ณ  ๋‹ค์Œ ์ƒ˜ํ”Œ์„ ์ฒ˜๋ฆฌํ•  ๋•Œ ์žฌ์‚ฌ์šฉํ•˜์ง€ ์•Š๋Š”๋‹ค. ์‹ ๊ฒฝ๋ง์ด ์ด์ „์— ์ฒ˜๋ฆฌํ–ˆ๋˜ ์ƒ˜ํ”Œ์„ ๋‹ค์Œ ์ƒ˜ํ”Œ์„ ์ฒ˜๋ฆฌํ•˜๋Š”๋ฐ ์žฌ์‚ฌ์šฉํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ด๋ ‡๊ฒŒ ๋ฐ์ดํ„ฐ ํ๋ฆ„์ด ์•ž์œผ๋กœ๋งŒ ์ „๋‹ฌ๋˜์–ด์„œ๋Š” ๊ณค๋ž€ํ•˜๋‹ค.๋‹ค์Œ ์ƒ˜ํ”Œ์„ ์œ„ํ•ด์„œ ์ด์ „ ๋ฐ์ดํ„ฐ๊ฐ€ ์‹ ๊ฒฝ๋ง ์ธต์— ์ˆœํ™˜๋  ํ•„์š”๊ฐ€ ์žˆ๋‹ค. ์ˆœํ™˜ ์‹ ๊ฒฝ๋ง(Recurrent Neural Network, RNN)์ˆœํ™˜ ์‹ ๊ฒฝ๋ง์—์„œ๋Š” ํŠน๋ณ„ํžˆ ์ธต์„ cell ์ด๋ผ๊ณ  ๋ถ€๋ฅธ๋‹ค. ํ•œ .. 2024. 11. 23.
[DL/WIL] 2 ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง(Convolution Neural Network, CNN) ํ•„ํ„ฐ์™€ ์ปค๋„Convolution Neural Network ์—์„œ๋Š” Fully Connected Neural Network ์™€๋Š” ๋‹ค๋ฅด๊ฒŒ ๋‰ด๋Ÿฐ์„ Filter ๋ผ๊ณ  ๋ถ€๋ฅธ๋‹ค.๋‰ด๋Ÿฐ = ํ•„ํ„ฐ = ์ปค๋„ ๋ชจ๋‘ ๊ฐ™์€ ๋ง์ด๋ผ๊ณ  ์ƒ๊ฐํ•ด๋„ ๋œ๋‹ค.Keras ์—์„œ Filter ๋Š” ๋‰ด๋Ÿฐ ๊ฐœ์ˆ˜๋ฅผ, Kernel ์€ ์ž…๋ ฅ์— ๊ณฑํ•ด์ง€๋Š” ๊ฐ€์ค‘์น˜๋ฅผ ์˜๋ฏธํ•œ๋‹ค. ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์—์„œ ์‹ค์ œ ๊ณ„์‚ฐ์€ ๋ฐ€์ง‘์ธต๊ณผ ๋™์ผํ•˜๊ฒŒ ๋‹จ์ˆœํžˆ ์ž…๋ ฅ๊ณผ ๊ฐ€์ค‘์น˜๋ฅผ ๊ณฑํ•˜๋Š” ๊ฒƒ์ด์ง€๋งŒ 2์ฐจ์› ํ˜•ํƒœ๋ฅผ ์œ ์ง€ํ•˜๋Š” ์ ์ด ๋‹ค๋ฅด๋‹ค. ๋˜ ์ž…๋ ฅ๋ณด๋‹ค ํ›จ์”ฌ ์ž‘์€ ํฌ๊ธฐ์˜ ์ปค๋„์„ ์‚ฌ์šฉํ•˜๊ณ  ์ž…๋ ฅ ์œ„๋ฅผ (์™ผ์ชฝ์—์„œ ์˜ค๋ฅธ์ชฝ์œผ๋กœ, ์œ„์—์„œ ์•„๋ž˜๋กœ) ์ด๋™ํ•˜๋ฉด์„œ 2์ฐจ์› ํŠน์„ฑ ๋งต์„ ๋งŒ๋“ ๋‹ค. ์ด๋ ‡๊ฒŒ 2์ฐจ์› ๊ตฌ์กฐ๋ฅผ ๊ทธ๋Œ€๋กœ ์‚ฌ์šฉํ•˜๊ธฐ ๋•Œ๋ฌธ์— ํ•ฉ์„ฑ๊ณฑ ์‹ ๊ฒฝ๋ง์ด ์ด๋ฏธ์ง€ ์ฒ˜๋ฆฌ ๋ถ„์•ผ์—์„œ ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ฐœํœ˜ํ•œ๋‹ค. ์ผ€๋ผ์Šค๋กœ ๊ตฌํ˜„ํ•œ ํ•ฉ์„ฑ๊ณฑ .. 2024. 11. 19.
[Paper Review] Pixel2Mesh: Generating 3D Mesh Models from Single RGB Images Pixel2Mesh: Generating 3D Mesh Models from Single RGB ImagesNanyang Wang1 โ‹†, Yinda Zhang2 โ‹†, Zhuwen Li3 โ‹†, Yanwei Fu4, Wei Liu5, Yu-Gang Jiang1 †1 Shanghai Key Lab of Intelligent Information Processing, School of Computer Science, Fudan University2 Princeton University 3 Intel Labs 4 School of Data Science, Fudan University 5 Tencent AI Lab Abstract.์ด ๋…ผ๋ฌธ์€ ๋‹จ์ผ ์ปฌ๋Ÿฌ ์ด๋ฏธ์ง€๋กœ๋ถ€ํ„ฐ 3D ํ˜•์ƒ์„ ์‚ผ๊ฐํ˜• ๋ฉ”์‰ฌ(triangular me.. 2024. 11. 6.
[Supplementary Review] PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation ์ด๋ฒˆ์—๋Š” " PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation " ๋…ผ๋ฌธ์˜ Supplementary ๋ฆฌ๋ทฐ๋กœ ๋Œ์•„์™”์Šต๋‹ˆ๋‹ค! Supplementary๋Š” ์ด B ~ H ๊นŒ์ง€ ์ด 7๊ฐœ์˜ ํŒŒํŠธ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.์—ฌ๊ธฐ์„œ ์ €๋Š” ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ชฉ์ฐจ๋กœ ๋ฆฌ๋ทฐํ•ด๋ณด๋ ค๊ณ  ํ•ฉ๋‹ˆ๋‹ค. [๋ชฉ์ฐจ]C : neural network architectures ์™€ training hyper parameters ์— ๋Œ€ํ•ด ๋””ํ…Œ์ผํ•˜๊ฒŒ ์•Œ์•„๋ณด๋Š” ์„น์…˜D : Detection pipeline ์— ๊ด€ํ•ด ๋””ํ…Œ์ผํ•˜๊ฒŒ ์•Œ์•„๋ณด๋Š” ์„น์…˜E : PointNet์˜ ์—ฌ๋Ÿฌ ์‘์šฉ ๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์•Œ์•„๋ณด๋Š” ์„น์…˜F : ๋”์šฑ ๋””ํ…Œ์ผํ•˜๊ฒŒ PointNet ๊ตฌ์กฐ๋ฅผ ๋ถ„์„ํ•ด๋ณด๋Š” ์„น์…˜  Section C : .. 2024. 11. 4.