HDR Scene Reconstruction from Resolution-Mismatched Event Streams and Single-Exposure Images

Area of Science:

• Computer Vision
• Computational Photography
• Robotics

Background:

• High dynamic range (HDR) reconstruction from low dynamic range (LDR) images is challenged by information loss in saturated areas.
• Event cameras offer high dynamic range but often have lower spatial resolution than RGB cameras, complicating hybrid sensor integration.
• Existing methods struggle with resolution mismatches, leading to color casts and artifacts when super-resolving event data for 3D radiance field reconstruction.

Purpose of the Study:

• To develop a framework for HDR radiance-field reconstruction from resolution-mismatched event and RGB camera inputs.
• To address the limitations of super-resolving event streams, such as global color cast and view-inconsistent artifacts.
• To improve the fidelity and visual quality of HDR reconstructions in hybrid sensor systems.

Main Methods:

• Proposed a framework integrating a pretrained 2D event super-resolution prior.
• Implemented a color correction module to anchor rendered radiance to LDR image chrominance.
• Introduced a dual-resolution event-stream constraint to supervise synthesized events at both super-resolved and native resolutions.

Main Results:

• Achieved higher six-scene mean HDR fidelity compared to the strongest baseline on the EvHDR-NeRF benchmark.
• Successfully reduced global color cast introduced by super-resolution preprocessing.
• Alleviated multi-view artifacts in the reconstructed radiance fields.

Conclusions:

• The proposed framework effectively handles resolution mismatches in hybrid event-image HDR reconstruction.
• The color correction and dual-resolution constraints are crucial for mitigating artifacts and improving fidelity.
• This method advances HDR radiance-field reconstruction for applications utilizing complementary sensor data.