A Novel Framework and Design Methodologies for Optimal Animation Production using Deep Learning
In this dissertation, we introduce an innovative automatic animation production framework and its modules aimed at overcoming the inherent challenges in traditional animation production, including the requirement for a substantial investment of time, resources, and expertise. The motivation for this scholarly endeavor stems from a comprehensive analysis of standard animation processes and consultations with experienced animators. Through this investigation, we have ascertained that achieving a high degree of automation necessitates the digitization of the full spectrum of resources and relevant data. To this end, we have structured the production continuum into four distinct but interrelated segments: Action List Generation, Stage Performance, Auto Cinematography, and Video Generation.In the realm of Stage Performance, we propose an innovative Real-time Phoneme Recognition Network (RealPRNet) for real-time lip-sync animation generation. Designed to correlate incoming audio input with the corresponding viseme (visual representation of a phoneme) in real-time, RealPRNet employs a multifaceted approach incorporating spatial, discrete temporal, and cohesive temporal features of the audio data. Significantly, the architecture of RealPRNet incorporates a stacked long short-term memory (LSTM) block and exploits long short-term features to optimize phoneme recognition. RealPRNet offers substantial improvements in reducing phoneme error rate and increasing in visual performance.Next, we delve into the text-to-animation framework (T2A), a pivotal construct in our research. While extant auto cinematography paradigms largely hinge on rudimentary cinematographic principles (aesthetics), T2A offers a more nuanced approach. It amalgamates the input script's visual fidelity with the resultant video's aesthetic compliance, by concurrently harnessing both fidelity and aesthetic models. Our evaluative studies, conducted with animators, attest to the superior quality of the resulting animations vis-à-vis methodologies solely reliant on aesthetic models, while reducing the workload of manual animation production by approximately 74%.However, it is germane to note that Auto Cinematography, when solely premised on canonical rule-based optimization, often falls short of audience expectations. we introduce RT2A (Reinforcement learning-based Text to Animation), a pioneering architecture that synergistically merges reinforcement learning techniques into the realm of automated film direction. Within the RT2A framework, each choice concerning camera orientation and positioning is systematically documented and integrated into subsequent training cycles for the reinforcement learning agent. Utilizing a carefully constructed reward function, the algorithm is guided toward the optimization of camera configurations that emulate the stylistic decisions commonly employed by human directors. Quantitative analysis reveals that RT2A can achieve 50% improvement in audience-approved camera placements and an 80% increase in fidelity over benchmark camera placements.Furthermore, due to the recent trend of personalized animation content influenced by user-generated interactions, the unpredictability and complexity of open-world settings pose a substantial challenge for automated cinematography techniques. To tackle this, we extend our framework along with an adaptive automated cinematography approach based on Generative Adversarial Networks (AACOGAN). According to empirical analyses, AACOGAN significantly outperforms existing methods in capturing the dynamism of open-world interactions. The efficacy of AACOGAN is underscored by a 73% improvement in the correlation between user behaviors and camera trajectory adaptations, as well as a marked elevation—up to 32.9%—in the quality of multi-focus scenes. These metrics offer compelling evidence that our methodology not only refines cinematographic outputs but also considerably augments user immersion within these complex virtual environments.
Read
- In Collections
-
Electronic Theses & Dissertations
- Copyright Status
- Attribution 4.0 International
- Material Type
-
Theses
- Authors
-
yu, zixiao
- Thesis Advisors
-
ren, jian
- Committee Members
-
li, tongtong
liu, sijia
li, wen
- Date
- 2023
- Subjects
-
Computer science
- Program of Study
-
Electrical Engineering - Doctor of Philosophy
- Degree Level
-
Doctoral
- Language
-
English
- Pages
- 153 pages
- Permalink
- https://doi.org/doi:10.25335/0twc-xv19