Reproduce “GARF: Gaussian Activated Radiance Fields for High Fidelity Reconstruction and Pose Estimation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, Simon Lucey

Gaussian Activated neural Radiance Fields (GARF) enable the joint optimization of 3D scene representation and camera poses by replacing traditional activations with Gaussian activations. This method eliminates the need for accurate prior poses and complex scheduling mechanisms, achieving superior view synthesis quality (e.g., PSNR 19.72 on the 'leaves' scene) and robust pose estimation (average rotation errors ~0.76 degrees) compared to previous approaches like BARF and SIREN.

Key insights

  • Gaussian activations enable smooth spectral control within the neural network, allowing continuous manipulation of the frequency response.
  • The use of Gaussian activations maintains accurate first-order derivatives, which is crucial for stable and robust pose optimization through gradient descent.
  • The continuous nature of Gaussian activations creates a smoother loss landscape, making the optimization less sensitive to weight initialization compared to sinusoidal networks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.