Novel view synthesis, a fundamental task in computer vision, aims to generate photo-realistic images from unseen viewpoints given a set of training images. Despite remarkable progress enabled by Neural Radiance Fields (NeRF), performance significantly...
Novel view synthesis, a fundamental task in computer vision, aims to generate photo-realistic images from unseen viewpoints given a set of training images. Despite remarkable progress enabled by Neural Radiance Fields (NeRF), performance significantly deteriorates under conditions where only a sparse set of training views is available, resulting in noisy artifacts, inaccurate geometry, and reduced visual fidelity. Recent approaches, including regularization-based and pre-training methods, have attempted to address these limitations but typically require extensive datasets or meticulously designed heuristics, often making them impractical for broader real-world applications.
In this dissertation, we propose advanced ray-processing strategies tailored explicitly for few-shot novel view synthesis, addressing shortcomings in existing methods. Specifically, we introduce three novel frameworks: MixNeRF, FlipNeRF, and ARC-NeRF, each targeting distinct but complementary aspects of ray-based scene modeling.
First, we present MixNeRF, which employs a mixture density model to represent rays. Unlike traditional approaches, MixNeRF robustly handles sparse inputs by effectively modeling complex multimodal distributions of blending weights along rays, thus leading to more precise 3D geometry under sparse-view conditions. This probabilistic representation significantly improves geometric consistency, reduces visual artifacts, and enhances overall rendering quality.
Second, we propose FlipNeRF, a novel ray augmentation technique that leverages estimated surface normals to generate flipped reflection rays. By introducing flipped reflection rays as additional training data, FlipNeRF enhances the accuracy of geometric reconstruction under sparse-view settings. To further refine training effectiveness, we introduce additional loss terms—Uncertainty-aware Emptiness Loss, which effectively mitigates floating artifacts by accounting for output uncertainty, and Bottleneck Feature Consistency Loss, which encourages feature-level consistency between paired rays, significantly boosting the fidelity and robustness of synthesized views.
Third, we introduce ARC-NeRF, an Area Ray Casting strategy that utilizes Integrated Positional Encoding (IPE) for improved ray parameterization. Unlike conventional ray augmentation methods, ARC-NeRF employs a single ray parameterized by IPE to cover broader unseen viewing regions effectively. Additionally, thanks to IPE-driven adaptive high-frequency regularization, ARC-NeRF dynamically regulates the high-frequency components of scene representation based on the pixel photo-consistency throughout the training, preventing early overfitting while preserving fine-grained detail.
Finally, we systematically analyze and integrate these advanced ray processing methods, demonstrating notable performance improvements over state-of-the-art baselines on benchmark datasets, including Realistic Synthetic 360°, DTU, LLFF, and Shiny Blender. Our comprehensive approach effectively addresses critical challenges inherent in sparse-view scenarios, advancing the practical applicability of NeRF-based novel view synthesis.