Work

IR-NeLF: infrared neural light field on mobile devices

Year: 2025

Type: article

Abstract: This paper proposes an infrared neural light field (IR-NeLF), which achieves infrared 3D reconstruction under challenging observation conditions (e.g., occlusion, poor visibility, etc.) while maintaining fast rendering performance on mobile devices. To address the challenge of matching using only infrared images, we design a strategy to recover the camera pose using only infrared images. This strategy first enhances the infrared images to overcome the inherent limitation of the indistinct detail features of infrared images in 3D reconstruction, and then restores the infrared camera pose through the traditional structure-from-motion algorithms. In addition, based on the low-frequency-dominant characteristics of infrared images, the network architecture is further optimized to reduce structural complexity, while a multi-dimensional hybrid loss function is incorporated to enhance the quality of image generation. Experimental results show that the proposed method achieves efficient rendering on edge devices, requiring only 53.6 ms to generate a 768 × 1024 infrared image on AIpro, significantly faster than MobileR2L (83.3 ms) and traditional NeRF (1360 s). In addition, it consistently outperforms the baselines in PSNR, SSIM, and LPIPS, demonstrating both efficiency and superior reconstruction quality. Finally, we contributed a high-resolution infrared image dataset to provide a foundation for future work.

Source: Optics Express

DOI: 10.1364/oe.574375

Authors: Mingjiang Ji, Yuanqing Wang, Xicai Li

Language: English

Topic: Advanced Vision and Imaging

Subfield: Computer Vision and Pattern Recognition

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $2,270

Funders: National Natural Science Foundation of China, National Key Research and Development Program of China

Awards: G3548313743, G4937485053