๋ณธ๋ฌธ ๋ฐ”๋กœ๊ฐ€๊ธฐ

์ „์ฒด ๊ธ€143

[Notable] Inference time vs Rendering time 1. Inference Time (์ถ”๋ก  ์‹œ๊ฐ„)"๋ชจ๋ธ์ด ์ž…๋ ฅ์„ ๋ฐ›์•„ ๊ฒฐ๊ณผ(์ถœ๋ ฅ)๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๊ฑธ๋ฆฌ๋Š” ์‹œ๊ฐ„"์˜ˆ: ์ด๋ฏธ์ง€ ํ•œ ์žฅ์„ ๋„ฃ์—ˆ์„ ๋•Œ, ๋ชจ๋ธ์ด 3D ์•„๋ฐ”ํƒ€๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๋ฐ ๊ฑธ๋ฆฌ๋Š” ์‹œ๊ฐ„.๋ณดํ†ต ๋”ฅ๋Ÿฌ๋‹ ๋„คํŠธ์›Œํฌ ๋‚ด๋ถ€ ์—ฐ์‚ฐ (forward pass)์— ํ•ด๋‹นํ•ด.์˜ˆ๋ฅผ ๋“ค์–ด, LHM ๋ชจ๋ธ์ด ๋‹จ์ผ ์ด๋ฏธ์ง€๋กœ๋ถ€ํ„ฐ 3D Gaussian ์•„๋ฐ”ํƒ€๋ฅผ ๋งŒ๋“ค์–ด๋‚ด๋Š” ๋ฐ ๊ฑธ๋ฆฌ๋Š” ์‹œ๊ฐ„์ด inference time๐Ÿ•’ ์‹œ๊ฐ„ ๋ฒ”์œ„: ์ˆ˜ ์ดˆ ~ ์ˆ˜ ๋ถ„๐Ÿ“ ํฌํ•จ ๋‚ด์šฉ:์ด๋ฏธ์ง€ ์ž…๋ ฅ → 3D representation ์˜ˆ์ธก๋„คํŠธ์›Œํฌ ์—ฐ์‚ฐ ๋ฐ post-processing 2. Rendering Time (๋ Œ๋”๋ง ์‹œ๊ฐ„)"์˜ˆ์ธก๋œ 3D ๋ฐ์ดํ„ฐ๋ฅผ ์‹œ๊ฐ์ ์œผ๋กœ ํ‘œํ˜„(๋ Œ๋”๋ง)ํ•˜๋Š” ๋ฐ ๊ฑธ๋ฆฌ๋Š” ์‹œ๊ฐ„"์˜ˆ: ์ƒ์„ฑ๋œ 3D Gaussian ์•„๋ฐ”ํƒ€๋ฅผ ํ™”๋ฉด์— ํฌํ† ๋ฆฌ์–ผ๋ฆฌ์Šคํ‹ฑํ•˜๊ฒŒ.. 2025. 3. 31.
[Notable] NeuS (Neural Implicit Surface) https://lingjie0206.github.io/papers/NeuS/ NeuS: Learning Neural Implicit Surfaces by Volume Rendering for Multi-view ReconstructionAbstract We present a novel neural surface reconstruction method, called NeuS (pronunciation: /nuหz/, same as "news"), for reconstructing objects and scenes with high fidelity from 2D image inputs. Existing neural surface reconstruction approaches, such alingjie0206.. 2025. 3. 18.
[Paper Review] SyncDreamer: Generating Multiview-Consistent Images From a Single-View Image https://arxiv.org/abs/2309.03453 SyncDreamer: Generating Multiview-consistent Images from a Single-view ImageIn this paper, we present a novel diffusion model called that generates multiview-consistent images from a single-view image. Using pretrained large-scale 2D diffusion models, recent work Zero123 demonstrates the ability to generate plausible novel views farxiv.org   Abstract์ด ๋…ผ๋ฌธ์—์„œ๋Š” SyncD.. 2025. 3. 18.
[Paper Review] Wonder3D: Single Image to 3D Using Cross-Domain Diffusion https://arxiv.org/abs/2310.15008 Wonder3D: Single Image to 3D using Cross-Domain DiffusionIn this work, we introduce Wonder3D, a novel method for efficiently generating high-fidelity textured meshes from single-view images.Recent methods based on Score Distillation Sampling (SDS) have shown the potential to recover 3D geometry from 2D diffusionarxiv.org ์ด ๋…ผ๋ฌธ์—์„œ ์ œ์‹œํ•˜๋Š” Wonder3D ๋Š” ๋‹ค์ค‘ ์‹œ์  ์ผ๊ด€์„ฑ์„ ์œ ์ง€ํ•˜๋Š” ๋ฒ•์„  .. 2025. 3. 13.
[Paper Review] ThemeStation: Generating Theme-Aware 3D assets from Few Exemplars https://arxiv.org/abs/2403.15383 ThemeStation: Generating Theme-Aware 3D Assets from Few ExemplarsReal-world applications often require a large gallery of 3D assets that share a consistent theme. While remarkable advances have been made in general 3D content creation from text or image, synthesizing customized 3D assets following the shared theme of inarxiv.org   ๋ฌธ์ œ ์ •์˜ ๊ฐ€์ƒํ˜„์‹ค(VR)์ด๋‚˜ ๋น„๋””์˜ค ๊ฒŒ์ž„์—์„œ๋Š” ํ…Œ๋งˆ์ ์œผ๋กœ.. 2025. 3. 11.
[Notable] ๋‹ค์ค‘๊ณต์„ ์„ฑ(Multicollinearity) ๋‹ค์ค‘๊ณต์„ ์„ฑ(Multicollinearity) ์ด๋ž€?: ํšŒ๊ท€ ๋ถ„์„์—์„œ ๋…๋ฆฝ ๋ณ€์ˆ˜๋“ค ๊ฐ„์— ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๊ฐ€ ์กด์žฌํ•˜๋Š” ๊ฒฝ์šฐ๋ฅผ ์˜๋ฏธํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์ค‘๊ณต์„ ์„ฑ์ด ์‹ฌํ•˜๋ฉด ํšŒ๊ท€ ๊ณ„์ˆ˜์˜ ์‹ ๋ขฐ์„ฑ์ด ๋‚ฎ์•„์ง€๊ณ , ์˜ˆ์ธก ๋ชจ๋ธ์˜ ํ•ด์„์ด ์–ด๋ ค์›Œ์ง„๋‹ค. ๋‹ค์ค‘๊ณต์„ ์„ฑ์ด ๋ฌธ์ œ๋ฅผ ์ผ์œผํ‚ค๋Š” ์ด์œ ํšŒ๊ท€ ๊ณ„์ˆ˜์˜ ๋ถˆ์•ˆ์ •์„ฑ๋…๋ฆฝ ๋ณ€์ˆ˜๋“ค์ด ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๊ฐ€์ง€๋ฉด, ํ•˜๋‚˜์˜ ๋ณ€์ˆ˜๋ฅผ ์กฐ๊ธˆ๋งŒ ๋ณ€๊ฒฝํ•ด๋„ ํšŒ๊ท€ ๊ณ„์ˆ˜๊ฐ€ ํฌ๊ฒŒ ๋ณ€ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.ํ•ด์„์˜ ์–ด๋ ค์›€์–ด๋–ค ๋…๋ฆฝ ๋ณ€์ˆ˜๊ฐ€ ์ข…์† ๋ณ€์ˆ˜์— ์‹ค์ œ๋กœ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š”์ง€ ํŒŒ์•…ํ•˜๊ธฐ ์–ด๋ ค์›Œ์ง‘๋‹ˆ๋‹ค.ํ†ต๊ณ„์  ์œ ์˜์„ฑ ๋ฌธ์ œํšŒ๊ท€ ๊ณ„์ˆ˜์˜ p-value ๊ฐ€ ๋†’์•„์ ธ ์œ ์˜ํ•˜์ง€ ์•Š๊ฒŒ ๋‚˜ํƒ€๋‚  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์ค‘๊ณต์„ ์„ฑ ์˜ˆ์‹œ1. ์‹ค์ƒํ™œ ์˜ˆ์‹œ: ๋ถ€๋™์‚ฐ ๊ฐ€๊ฒฉ ์˜ˆ์ธก์•„ํŒŒํŠธ ๊ฐ€๊ฒฉ์„ ์˜ˆ์ธกํ•˜๋Š” ํšŒ๊ท€ ๋ชจ๋ธ์„ ๋งŒ๋“ ๋‹ค๊ณ  ๊ฐ€์ •ํ•ด ๋ด…์‹œ๋‹ค. ๋…๋ฆฝ ๋ณ€์ˆ˜๋กœ ๋‹ค์Œ์„ ํฌํ•จํ•œ๋‹ค๊ณ  ๊ฐ€์ •ํ•ฉ๋‹ˆ๋‹ค.์ „์šฉ๋ฉด์  .. 2025. 3. 4.