Co-Adaptive Neural Radiance Field With Geometry-Feature

Recent 3D reconstruction methods based on Neural Radiance Fields (NeRF) have significantly improved the anti-aliasing effect and training speed through implicit neural representations and acceleration frameworks. However, these methods generally adopt fixed-frequency positional encoding, which is difficult to adapt to the texture complexity of different regions, resulting in an imbalance in the frequency allocation of high-frequency/low-frequency signals. We propose a hybrid framework combining geometric decomposition and featureaugmented positional encoding to deal with the problem. Our approach disentangles 3D coordinates into orthogonal sub-planes using Multi-Layer Perception (MLP) with attention-weighted fusion, reducing spatial entanglement. And the feature-augmented positional encoding dynamically adjusts frequency allocation based on local texture density using learnable 2D projections. Hierarchical sampling further prioritizes rays with high spatialfrequency gradients. Experiments demonstrate improved reconstruction of structured geometries and high-frequency details, although feature stability requires regularization. This work bridges explicit geometric priors with data-driven adaptation.