Research map / Architectures and efficiency / Spiking networks
Spiking NeRF: Representing the Real-World Geometry by a Discontinuous Representation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Spiking NeRF, which replaces the density activation with a bounded full-precision integrate-and-fire spiking neuron (B-FIF) in a hybrid ANN-SNN framework to build a discontinuous density field for more accurate surface geometry.
Problem
Continuous MLP-based density fields in NeRF cannot represent the discontinuity at air-surface interfaces, causing threshold-dependent post-processing, view-dependent optimal thresholds, and disappearing geometry for semi-transparent or thin objects.
Approach
Replace the final density activation with a bounded full-precision integrate-and-fire neuron whose learnable threshold and maximum activation (k·tanh with learnable r and k) are tied by a derived depth-error bound, so the spiking threshold acts as a learned surface threshold while a single time step keeps training ANN-like.
Key results
- On the Blender dataset, average Chamfer distance 1.07 (x10^-2) vs. 1.09 for PET-NeuS and 1.64 for Mip-NeRF, the lowest average.
- On light-density scenes (Dex-NeRF and self-created Blender scenes), average Chamfer distance 8.75 (x10^-2) vs. 9.77 for NeRF and 10.16 for Mip-NeRF.
- On DTU, average Chamfer distance 0.90, tied with HF-NeuS and behind PET-NeuS (0.83), which the authors attribute to SDF-based methods' strengths on smooth scenes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a discontinuous density representation with B-FIF neurons | Spiking NeRF · Dex-NeRF and Blender light scenarios (Avg.) | Chamfer distance × 10^-2: 9.77 -> 8.75 | Helped |
| Use a discontinuous density representation with B-FIF neurons | Spiking NeRF · Blender (Avg.) | Chamfer distance × 10^-2: 1.09 -> 1.07 | Helped |
| Use a discontinuous density representation with B-FIF neurons | Spiking NeRF · DTU (Avg.) | Chamfer distance: 0.83 -> 0.90 | Hurt |
| Add the spiking-threshold regularization L_v (ablation) | Spiking NeRF · surface reconstruction | Chamfer distance | Helped |
| Constrain maximum activation with B-FIF instead of using unbounded FIF (ablation) | Spiking NeRF · surface reconstruction | Chamfer distance | Helped |
| Add the initial smoothness loss L_g (ablation) | Spiking NeRF · surface reconstruction | Chamfer distance | No clear effect |
Surprising findings
- The authors note that it is rare for an SNN to outperform an ANN, and their SNN-based NeRF beats ANN-based baselines on several benchmarks.
- Directly using IF or FIF neurons in the density head gives large depth error or poor accuracy, so the bounded B-FIF variant is needed.
Limitations
- Acknowledged gap between the theoretical depth-error bound and actual average depth error, so the bound is not yet practical for ground-truth-free evaluation.
- Does not specifically address scenes with high light and low brightness, which may still cause geometry errors.
- On DTU, results are less competitive than PET-NeuS and HF-NeuS in some scenes.
Experimental setup
Datasets: Blender Dex-NeRF DTU
Benchmarks: Blender Dex-NeRF DTU
Metrics: Chamfer distance depth error
Baselines: NeRF Mip-NeRF NeuS HF-NeuS PET-NeuS
Builds on: NeRF NeuS Full-precision integrate-and-fire (FIF) NeRF++ Mip-NeRF
Closest papers by meaning
- Spik-NeRF: Spiking Neural Networks for Neural Radiance Fields (NeurIPS 2025 · similarity 0.88)
- Spiking Graph Neural Network on Riemannian Manifolds (NeurIPS 2024 · similarity 0.85)
- Take A Shortcut Back: Mitigating the Gradient Vanishing for Training Spiking Neural Networks (NeurIPS 2024 · similarity 0.85)
- Implicit-ARAP: Efficient Handle-Guided Neural Field Deformation via Local Patch Meshing (NeurIPS 2025 · similarity 0.84)
- SpikeNeRF: Learning Neural Radiance Fields from Continuous Spike Stream (CVPR 2024 · similarity 0.84)
- Sign Gradient Descent-based Neuronal Dynamics: ANN-to-SNN Conversion Beyond ReLU Network (ICML 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.