Abstract
Neural Radiance Fields (NeRFs) utilizing multiresolution hashgrids, such as Instant-NGP, offer rapid training but suffer from hash collisions and high memory usage. HollowNeRF addresses these issues by learning to "hollow out" solid objects, pruning invisible internal features that contribute to collisions. By integrating a trainable 3D saliency grid, a zero-skipping gate, and an ADMM-based pruner, HollowNeRF reduces parameter counts by up to 69% while maintaining or improving rendering quality compared to state-of-the-art baselines.
Introduction
The advent of Instant-NGP introduced the multiresolution hashgrid, significantly accelerating NeRF training and inference. However, this approach is limited by hash collisions, where multiple 3D coordinates map to the same feature table entry, leading to reconstruction noise. Furthermore, standard NeRFs redundantly represent the interior of solid objects, wasting capacity on invisible regions. HollowNeRF proposes a learned sparsity framework that identifies and prunes these unimportant regions, effectively "hollowing" objects to reclaim hashgrid capacity and mitigate interference.
Methodology
HollowNeRF introduces three primary architectural components to achieve learned sparsity and collision mitigation:
- Trainable 3D Saliency Grid ($G$): A coarse grid assigns saliency weights to spatial regions. Features fetched from the hashgrid are modulated by these weights; a weight of zero effectively prunes the feature.
- Soft Zero-Skipping Gate: To ensure that pruned features result in zero spatial density ($\sigma$), a $\tanh$-based gate is added to the decoder. This prevents the model from hallucinating density in regions where features have been zeroed out.
- ADMM Pruner: The Alternating Direction Method of Multipliers (ADMM) is employed to enforce sparsity constraints on the saliency grid. This optimization pushes the model to identify the minimal set of voxels (the "shell") required for accurate rendering.
By zeroing out features for internal, invisible voxels, HollowNeRF prevents these voxels from "polluting" shared hash table entries, which directly improves the signal-to-noise ratio for visible surface voxels.
Results and Discussion
Evaluations on the NeRF Synthetic and Tanks and Temples datasets demonstrate significant efficiency gains. HollowNeRF achieves rendering quality comparable to Instant-NGP while utilizing only 31% of the parameters. When parameter counts are increased to approximately 56% of the baseline, HollowNeRF provides a 1dB PSNR improvement, indicating that collision mitigation directly translates to higher visual fidelity and sharper details.
Qualitative analysis reveals that HollowNeRF successfully captures fine textures and geometric details that are typically blurred by collisions in standard hashgrid models. The "hollowing" effect is clearly visible in cross-sections, where the model focuses its capacity on the object's surface.
Conclusion
HollowNeRF demonstrates that learned sparsity is a powerful tool for optimizing hashgrid-based NeRFs. By hollowing out objects, the method simultaneously reduces memory footprint and improves rendering quality through collision mitigation. While highly effective for sparse scenes, future work may explore extending this framework to handle semi-transparent media and complex specular reflections.
Summary updated