BEGIN:VCALENDAR VERSION:2.0 PRODID:Linklings LLC BEGIN:VTIMEZONE TZID:Asia/Tokyo X-LIC-LOCATION:Asia/Tokyo BEGIN:STANDARD TZOFFSETFROM:+0900 TZOFFSETTO:+0900 TZNAME:JST DTSTART:18871231T000000 END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20250110T023312Z LOCATION:Hall B5 (2)\, B Block\, Level 5 DTSTART;TZID=Asia/Tokyo:20241205T165300 DTEND;TZID=Asia/Tokyo:20241205T170500 UID:siggraphasia_SIGGRAPH Asia 2024_sess137_papers_887@linklings.com SUMMARY:StableNormal: Reducing Diffusion Variance for Stable and Sharp Nor mal DESCRIPTION:Technical Papers\n\nChongjie Ye and Lingteng Qiu (FNii, The Ch inese University of Hong Kong, Shenzhen; SSE, The Chinese University of Ho ng Kong, Shenzhen); Xiaodong Gu and Qi Zuo (Alibaba); Yushuang Wu (FNii, T he Chinese University of Hong Kong, Shenzhen; SSE, The Chinese University of Hong Kong, Shenzhen); Zilong Dong and Liefeng Bo (Alibaba); Yuliang Xiu (Max Planck Institute for Intelligent Systems); and Xiaoguang Han (SSE, T he Chinese University of Hong Kong, Shenzhen; FNii, The Chinese University of Hong Kong, Shenzhen)\n\nThis work addresses the challenge of high-qual ity surface normal estimation from monocular colored inputs (i.e., images and videos), a field which has recently been revolutionized by repurposing diffusion priors. However, previous attempts still struggle with stochast ic inference, conflicting with the deterministic nature of the Image2Norma l task, and costly ensembling step, which slows down the estimation proces s. Our method, StableNormal, mitigates the stochasticity of the diffusion process by reducing inference variance, thus producing “Stable-and-Sharp” normal estimates without any additional ensembling process. StableNormal w orks robustly under chal lenging imaging conditions, such as extreme light ing, blurring, and low quality. It is also robust against transparent and reflective surfaces, as well as cluttered scenes with numerous objects. Sp ecifically, StableNormal employs a coarse-to-fine strategy, which starts w ith a one-step normal estimator (YOSO) to derive an initial normal guess, that is relatively coarse but reliable, then followed by a semantic-guided refinement process (SG-DRN) that refines the normals to recover geometric details. The effectiveness of StableNormal is demonstrated through compet itive performance in standard datasets such as DIODE-indoor, iBims, Scanne tV2, and NYUv2, and also in various downstream tasks, such as surface reco nstruction and normal enhancement. These results evidence that StableNorma l retains both the “stability” and “sharpness” for accurate normal estimat ion. StableNormal represents a baby attempt to repurpose diffusion priors for deterministic estimation. To democratize this, code and models have be en publicly available in hf.co/Stable-X.\n\nRegistration Category: Full Ac cess, Full Access Supporter\n\nLanguage Format: English Language\n\nSessio n Chair: Michael Rubinstein (Google) URL:https://asia.siggraph.org/2024/program/?id=papers_887&sess=sess137 END:VEVENT END:VCALENDAR