[Paper] EventNeuS: 3D Mesh Reconstruction from a Single Event Camera

Source: arXiv - 2602.03847v1

Overview

EventNeuS introduces a self‑supervised neural pipeline that can reconstruct high‑quality 3D meshes from the sparse, asynchronous data of a single event camera. By marrying signed distance function (SDF) learning with event‑based supervision, the method pushes the accuracy of event‑driven 3‑D reconstruction well beyond prior work, opening the door for lightweight, low‑latency perception in robotics and AR/VR.

Key Contributions

• First event‑camera pipeline that learns a continuous SDF + density field for full 3‑D mesh recovery from a monocular event stream.
• Self‑supervised training that requires no ground‑truth geometry or depth maps; the model is driven solely by the event data itself.
• Spherical‑harmonics encoding integrated into the neural representation to capture view‑dependent lighting and motion blur effects that are common in event streams.
• Significant quantitative gains: 34 % lower Chamfer distance and 31 % lower mean absolute error (MAE) compared with the strongest existing event‑based reconstruction baseline.
• Open‑source implementation (code and pretrained weights) that can be plugged into existing event‑camera pipelines.

Methodology

1. Event Stream Pre‑processing – The raw asynchronous events (x, y, t, polarity) are accumulated into short, overlapping time windows, producing a set of “event frames” that retain high temporal resolution while providing enough density for neural processing.

2. Neural Implicit Representation – A multi‑layer perceptron (MLP) predicts two fields for any 3‑D point:

   • Signed Distance Function (SDF) – encodes the surface geometry.
   • Density (σ) – models volumetric opacity, enabling differentiable rendering of the event‑camera’s motion‑blurred measurements.

3. Spherical Harmonics Lighting – Instead of a simple color/texture branch, the network learns a low‑order spherical‑harmonics basis that modulates the emitted intensity based on viewing direction. This handles the strong view‑dependent contrast changes typical of event cameras.

4. Self‑Supervised Event Loss – The predicted density field is rendered into synthetic event frames using a differentiable event generation model (event polarity is derived from temporal intensity gradients). The loss compares these synthetic events with the real ones via a combination of binary cross‑entropy (polarity) and L1 (timestamp) terms.

5. Optimization – The whole system is trained end‑to‑end with Adam, alternating between geometry updates (SDF) and appearance updates (spherical harmonics). No external supervision (e.g., depth sensors) is required.

Results & Findings

• Geometry quality: The reconstructed meshes capture fine details (e.g., thin edges, small protrusions) that previous event‑based methods missed.
• Robustness to motion: Because the density field models motion blur, the system remains stable even when the camera or scene moves rapidly.
• Generalization: Experiments on indoor tabletop objects, outdoor street scenes, and fast‑moving drones show consistent improvements without re‑training per domain.

Practical Implications

• Low‑Power Robotics – Event cameras consume orders of magnitude less power than RGB‑D sensors. EventNeuS enables robots to build dense 3‑D maps on the edge, ideal for drones, micro‑robots, or wearables where battery life is critical.
• High‑Speed SLAM – The method’s ability to handle rapid motion without motion‑blur artifacts makes it a strong candidate for SLAM pipelines in high‑speed navigation (e.g., autonomous racing).
• AR/VR Headsets – Event‑based depth estimation can complement eye‑tracking and foveated rendering, providing real‑time scene geometry without the latency of traditional depth cameras.
• Industrial Inspection – In environments with harsh lighting (e.g., welding, night‑time inspection), event cameras thrive; EventNeuS can generate accurate CAD‑ready meshes for defect detection.

Developers can integrate the provided PyTorch module into existing perception stacks, swapping out the RGB‑based mesh reconstruction component with an event‑driven one that runs on modest GPUs or even embedded AI accelerators.

Limitations & Future Work

• Static‑Scene Assumption for SDF – The current formulation assumes a mostly static geometry; moving objects are encoded as noise in the density field. Extending to dynamic SDFs would enable full scene flow reconstruction.
• Resolution Trade‑off – While the method works well on medium‑scale objects, reconstructing large outdoor environments still requires hierarchical or multi‑scale strategies to keep memory usage tractable.
• Event Noise Sensitivity – In extremely low‑light conditions, event sparsity can degrade the self‑supervised loss; incorporating a denoising front‑end or hybrid RGB‑event fusion could mitigate this.

Future research directions include dynamic scene modeling, multi‑camera event fusion, and real‑time integration with SLAM back‑ends for closed‑loop navigation.

Authors

• Shreyas Sachan
• Viktor Rudnev
• Mohamed Elgharib
• Christian Theobalt
• Vladislav Golyanik

Paper Information

• arXiv ID: 2602.03847v1
• Categories: cs.CV
• Published: February 3, 2026
• PDF: Download PDF