Borom Tunwattanapong's research while affiliated with Institute For Creative Technologies and other places

... With the objective of harnessing multiple dimensions of light concurrently, recent methods have utilized both polarization and spatial dimensions. These have found applications such as 3D reconstruction with a polarized light stage [Ghosh et al. 2011] and multi-view capture with a hand-held polarization camera [Dave et al. 2022;Hwang et al. 2022]. Image segmentation leveraging polarization, spectrum, and spatial dimensions was demonstrated by Chen et al. [2022]. ...

... Although the virtual light stage provides a way to learn the physically-accurate relighting behaviors from synthetic data, the realism of the deep network output is often limited when trained with only synthetic data. After all, the photogrammetry face scan comes with only basic reflectance information such as diffuse albedo and does not provide all the necessary information for realistic portrait relighting, such as surface [Graham et al. 2013;Nagano et al. 2015] and subsurface [Jensen et al. 2001] reflectance properties. As a result, the generated synthetic training data exhibits a domain gap from their real counterparts. ...

... Fine detail recorded through capture technology[30]. ...

... Photometric Stereo (PS) approaches [27], [28], [29], [30], [31], [32], [33] assume three or more images captured with a static camera while varying illumination or object pose [34], [35] to resolve the aforementioned ambiguities. In contrast to early PS approaches which often assumed orthographic cameras and distant light sources, newer works have considered the more practical setup of near light sources [36], [37], [38], [39] and perspective projection [40], [41], [42]. ...

... (1) IVAs have transitioned from the lab to the real world, indicating increased technical maturity and robustness [4,20,29,33] (2) The prevalence of digital personal assistants has resulted in core IVA-related commercial technologies increasing in availability and quality (e.g., ASR, NLP, TTS) [10,48] (3) Computing has moved from predominantly desktop-centric to various hardware platforms (e.g., web, mobile, XR) [11,16] (4) The cloud has been embraced by academia and industry, leading to more runtime and distribution options [8,15] (5) The visual fidelity and availability of character art assets has continued to increase [9,12,50] These trends point towards the need to re-architect the platform to a VHToolkit 2.0 in order to 1) take advantage of modern architecture and software development principles, 2) leverage commercial offerings relevant to IVAs, 3) facilitate cross-pollination with other academic efforts, and 4) support multiple hardware platforms. This paper describes the approach and current status of our ongoing work in support of these goals together with lessons learned during the re-architecting process. ...

... Dong et al. [2015] used an interferometry device to acquire the microstructure of brushed metal, but they still use statistical reflectance models to fit the measured data for rendering. Other methods [Graham et al. 2013;Nagano et al. 2015;Nam et al. 2016] aim at measuring accurate heightfields; these could be used with glint rendering methods, but seamless extension of the data across larger surface areas remains a problem, which could be addressed by our method. ...

... Other methods [Zhou and Kambhamettu 2002;Matusik et al. 2003;Zickler et al. 2006;Xu and Wallace 2008;Tunwattanapong et al. 2011;Lombardi and Nishino 2012;Ren et al. 2015] require multiple input images or special equipment while capturing the image to collect the data required to estimate illumination. Our method differs from these in that it only needs a single LDR image and a small amount of user annotations and no access to geometry or any devices other than a digital camera. ...

... This approach applied to reconstruction of 3D faces from unknown orientation and restoring missing surface regions of face. This approach only generates 3D faces of known face images and constructs missing part of surface.Besides 3D face modeling techniques T. Weise et al.[11,12]has been given approach to manage facial poses and expressions based on live performance of discrimatory or choosy key facial points. These approaches controlls low value 2D images and 3D depth maps to realism based expressions.This system has drawback of it requires human faces as well as cartoon faces which are widely used to generate animated character.E. ...

... Texture alignment to synthesize parametric textures for a 3D model has been recently tackled by using online optical flow computed from the virtual view-point [29]. In addition, a combination of image, shape and directable forces has been used to create scan correspondences in [30] . Similarly, multicamera setups also use texture alignment techniques to synthesize blended view-dependent appearances [31, 32]. ...

... Despite considerable progress in this field, there are still many challenges to overcome, including expensive and sophisticated setup for avatar capture, lack of support for realistic relighting and animation, and high resource demands making training time-consuming and realtime deployment difficult to achieve. Traditional frameworks based on graphics pipeline, including geometry reconstruction [Beeler et al. 2010[Beeler et al. , 2011Collet et al. 2015;Guo et al. 2019;Riviere et al. 2020;] and physically-inspired reflectance capture [Debevec et al. 2000;Ghosh et al. 2011;Ma et al. 2007;Moser et al. 2021;Weyrich et al. 2006], are often difficult to set up and lack robustness, especially for dynamic subjects and non-facial parts. Recent deep learning based methods [Bi et al. 2021;Lombardi et al. 2018Remelli et al. 2022] have demonstrated promising improvements for avatar representation by approximating the geometry and appearance with neural networks. ...