Problem
Research questions and friction points this paper is trying to address.
Mixed Neural Fields
Real-world Data Accuracy
Image Reconstruction Error
Canonical Factors for Hybrid Neural Fields
🤖 AI Summary
Hybrid neural fields exhibit substantial reconstruction bias (up to 2 PSNR) and low efficiency on axis-aligned real-world data, primarily due to systematic distortions introduced by factorized feature volumes.
Method: We propose learning a normalized coordinate transformation to eliminate this bias, enabling an efficient and interpretable hybrid neural field. We theoretically establish and empirically validate the feasibility and benefit of jointly optimizing the coordinate transform and scene representation. We further design the TILTED architecture, which—through controlled ablation—first exposes systemic flaws in standard evaluation protocols for neural field modeling.
Results: Our method achieves significant improvements over baselines across image, signed distance function (SDF), and NeRF reconstruction tasks: +PSNR, enhanced robustness, 2× model compression, accelerated inference, and performance on par with models twice its parameter count.
📝 Abstract
Factored feature volumes offer a simple way to build more compact, efficient, and intepretable neural fields, but also introduce biases that are not necessarily beneficial for realworld data. In this work, we (1) characterize the undesirable biases that these architectures have for axis-aligned signals—they can lead to radiance field reconstruction differences of as high as 2 PSNR—and (2) explore how learning a set of canonicalizing transformations can improve representations by removing these biases. We prove in a simple two-dimensional model problem that a hybrid architecture that simultaneously learns these transformations together with scene appearance succeeds with drastically improved efficiency. We validate the resulting architectures, which we call TILTED, using 2D image, signed distance field, and radiance field reconstruction tasks, where we observe improvements across quality, robustness, compactness, and runtime. Results demonstrate that TILTED can enable capabilities comparable to baselines that are 2x larger, while highlighting weaknesses of standard procedures for evaluating neural field representations.
Innovation
Methods, ideas, or system contributions that make the work stand out.
TILTED method
feature volume accuracy
enhanced model efficiency
🔎 Similar Papers
No similar papers found.
