Neural Network Inductive Biases for High-fidelity 3D Reconstruction

High-fidelity 3D reconstruction demands representations that capture both coarse structure and subtle geometric details. Implicit Neural Representations can encode 3D geometry at arbitrary resolution, yet their accuracy is limited by the inductive biases of coordinate multilayer perceptrons, which favor low-frequency over high-frequency content.

I analyzed inductive biases for ReLU-based and sine-based implicit neural representations and proposed learnable sinusoidal activations that extend fixed-frequency SIREN layers.

By letting the network tune amplitudes, frequencies, and phase shift of its activation functions depending on the input coordinate, the network can recover finer details and achieve lower reconstruction error relative to SIREN at a comparable parameter count, while also being more robust to the initial hyperparameters.