Single Image 3D Hair Modeling Techniques And Applications

Hair is crucial for the perceived identity of a human being, and often reveals mul-tiple aspects of the subject. Due to the extreme geometric complexity, image-based hair modeling has long been a challenging task. Manual hair modeling methods require professional skills and tedious labor work, while existing hair acquisition methods only work on multi-view input captured in controlled environment and with special equip-ments. Therefore, these techniques are not suitable for consumer-level practices.In this dissertation, we propose a series of novel single-view hair modeling tech-niques, which are made possible by combining minimal user-assistance, machine learn-ing methods and geometry priors. Meanwhile, in consideration of the prohibitively ex-pensive computation cost of hair animation, we also propose an efficient hair simulation method based on hair skinning model, which is able to achieve high-quality hair ani-mation in real-time and handle detailed hair-solid interactions. We fully demonstrate the effectiveness of these techniques with various intelligent portrait editing applications-Specifically, this dissertation mainly includes the following five parts:1. Focusing on portrait hair manipulation, we first propose an interactive single-view hair modeling method. It is able to generate a high-resolution strand-based hair model with limited user-input, that can be faithfully re-rendered in novel view-points, to support various applications such as portrait pop-up and hairstyle transfer.2. Based on the interactive method, we propose an automatic single-view hair model-ing method, which removes user-assistance entirely without sacrificing the quality.This helps to popularize related techniques and enable large-scale hair generation.3. Aiming at the geometric quality, we propose a method for single-view high-quality portrait model reconstruction. This method combines multiple image-based and prior-based clues jointly to generate highly-detailed portrait surface for related applications, such as portrait relighting and portrait relief printing.4. As for high-performance hair simulation, we propose an efficient data-driven hair simulation framework based on hair skinning model, which is able to learn fun-damental hair mutual motion relationship and drive speedy run-time simulation.Meanwhile, it is also equipped with strand collision correction to recover motion details lost due to interpolation.5. Built upon the data-driven simulation framework, we propose to improve it with adaptive skinning model to better resolve hair-solid interaction, which violates the assumption of the previous reduced model. Consequently, it is able to perform detailed interactive hair-solid simulation without affecting the performance.