A Survey of 3D Reconstruction with Event Cameras

A comprehensive survey reviews 3D reconstruction techniques using event cameras, which capture asynchronous per-pixel brightness changes rather than traditional frames. The research categorizes methods across stereo, monocular, and multimodal systems using geometry-based, deep learning, and neural rendering approaches, identifying key challenges in datasets, evaluation standards, and dynamic scene handling.

Event cameras represent a paradigm shift in machine vision, offering fundamentally different data acquisition compared to conventional frame-based sensors. By capturing brightness changes asynchronously at the pixel level, these devices generate sparse yet temporally dense information streams, enabling reconstruction in scenarios where traditional cameras struggle—high-speed motion, low-light environments, and extreme dynamic range situations. This survey synthesizes the rapidly evolving field of event-based 3D reconstruction, providing systematic categorization of approaches across input modalities and methodologies.

The convergence of event cameras with advanced computational techniques marks a significant advancement in computer vision. Geometry-based approaches leverage traditional photogrammetry principles adapted for sparse event data, while deep learning methods extract features directly from event streams without explicit geometric modeling. Emerging neural rendering techniques like NeRF and 3D Gaussian Splatting demonstrate how event data integrates with implicit scene representations, opening new possibilities for efficient and accurate reconstruction.

The practical implications extend across robotics, autonomous vehicles, and aerial systems where real-time performance and reliability under adverse conditions determine operational success. Autonomous driving systems benefit from event cameras' superior temporal resolution during rapid scene changes and low-visibility conditions. The robotics and aerial navigation sectors gain enhanced spatial awareness capabilities critical for navigation and obstacle avoidance.

Critical bottlenecks remain in dataset standardization, evaluation metrics, and theoretical understanding of event representation. Future progress depends on establishing benchmark datasets comparable to traditional vision benchmarks, developing standardized evaluation protocols, and advancing techniques for dynamic scene reconstruction where object motion complicates 3D recovery. These improvements will accelerate practical deployment across industrial applications.

• →Event cameras enable 3D reconstruction in challenging conditions like high-speed motion and low illumination where frame-based cameras fail.
• →Current approaches span geometry-based methods, deep learning, and neural rendering techniques including NeRF and 3D Gaussian Splatting.
• →Key limitations include insufficient public datasets, lack of standardized evaluation metrics, and difficulties reconstructing dynamic scenes.
• →Applications span autonomous driving, robotics, aerial navigation, and virtual reality with significant real-time performance advantages.
• →Future research must prioritize dataset standardization and evaluation protocols to accelerate commercial deployment across industries.