LarsNeRF: Fast and Effective View Synthesis From Low-Altitude Remote Sensing Scenarios

Zhong Chen, Siyang Wang, Liu Liu

Abstract: The neural radiance field (NeRF) has garnered significant attention for its ability to synthesize photorealistic views, when applied to low-altitude remote sensing scenes, NeRF encounters three challenges: 1) low synthesis efficiency due to extreme scale changes in scenes; 2) blurred details resulting from difficulty in balancing far and near perspectives; and 3) glimmering artifacts due to transient object occurrences. To address these issues, we propose a fast and effective view synthesis approach named LarsNeRF, consisting of three key components: regularized multiresolution hash encoding (to enhance efficiency), multiscale feature enhancement (to restore details), and antitransient interference (to eliminate artifacts). Experimental results on both real and simulated datasets demonstrate that LarsNeRF not only boosts efficiency by 30 times but also improves average PSNR by 2 dB compared to the baseline NeRF. Additionally, LarsNeRF has an efficiency improvement of nearly three times compared to Mip-NeRF, and the image quality is almost the same. LarsNeRF exhibits significant advantages in view synthesis from low-altitude remote sensing scenes compared to state-of-the-art approaches. These findings indicate that LarsNeRF is suitable for real-time and effective view synthesis in low-altitude remote sensing scenes.