Abstract: Scene reconstruction of hutong-style blocks plays a vital role in promoting and preserving the cultural heritage of China's historic streets. However, the unique characteristics of hutong-style neighborhoods, such as their narrow and elongated layouts, significant variations in imaging distances, and differences in resolution, result in suboptimal reconstruction outcomes. To address these challenges, this paper proposes a 3D reconstruction algorithm for narrow blocks based on neural radiation field (NeRF), capable of reconstructing high-quality scenes under the condition of outdoor panoramic images. To solve the problem of uneven panoramic image, the algorithm introduces a distortion-aware ray sampling method, which adjusts the sampling frequency to enhance model learning quality. Additionally, to address the narrowness of scenes in the hutong area, a direction-sensitive color-depth-structure joint loss function is constructed to improve the adaptability of the model to the narrow block and thereby enhancing reconstruction quality. Experimental results demonstrate that, compared to baseline methods that use distortion-aware ray sampling only, the proposed approach achieves higher peak signal-to-noise ratio (PSNR) values and delivers high-quality novel view synthesis.

Key words: hutong-style blocks, NeRF, 3D reconstruction, panoramic image, new perspective synthesis