This dissertation presents structure-aware methods for adaptive density control in 3D Gaussian Splatting (3DGS) and deep learning-based camera parameter estimation from single-view images. The methods exploit spatial structures that are informative for each problem: local structural complexity and insufficiently represented spatial details for 3D Gaussian refinement, and spatially consistent object edges as implicit geometric cues for camera parameter estimation. In 3DGS, standard adaptive density control (ADC) mainly relies on positional-gradient magnitudes, and can therefore miss fine details, redundantly split primitives that do not require further subdivision, or place sub-primitives without considering local structure. To better exploit positional gradients, we introduce directional consistency-driven density control, which measures the angular coherence of gradients within each primitive and uses it for both split selection and split placement. This measure captures local structural complexity and guides both whether and where a primitive is split, reducing primitive counts by up to 30% while improving reconstruction quality. We further introduce blur-driven density control, which uses differentiable per-primitive blur to identify regions where spatial details remain insufficiently represented and can be missed by positional-gradient-based ADC. The learned blur levels guide selective split and clone operations toward these regions, while detail-aware regularization and blur-adaptive opacity attenuation suppress redundant primitives in smooth regions or low-contribution primitives. This improves fine-detail reconstruction while reducing primitive counts by up to 42% over 3DGS-based baselines. For camera parameter estimation, we present a deep learning-based weighted edge-attention model that estimates camera rotations, field of view, and distortion from single-view images. Unlike classical methods constrained by explicit geometric cues such as vanishing points, the model employs object edges as implicit geometric cues that can be observed more frequently in images. The model emphasizes edges from semantic classes with stable size and shape statistics, because spatially consistent object edges better reflect geometric appearance changes caused by camera parameters. The model reduces RMSE by up to 80% over previous learning-based methods in our benchmark.