中国科学院自动化研究所机构知识库

Neural Radiance Field (NeRF) can render complex 3D scenes with viewpoint-dependent effects. However, less work has been devoted to exploring its limitations in high-resolution environments, especially when upscaled to ultra-high resolution (e.g., 4k). Specifically, existing NeRF-based methods face severe limitations in reconstructing high-resolution real scenes, for example, a large number of parameters, misalignment of the input data, and over-smoothing of details. In this paper, we present a novel and effective framework, called De-NeRF, based on NeRF and deformable convolutional network, to achieve high-fidelity view synthesis in ultra-high resolution scenes: (1) marrying the deformable convolution unit which can solve the problem of misaligned input of the high-resolution data. (2) Presenting a density sparse voxel-based approach which can greatly reduce the training time while rendering results with higher accuracy. Compared to existing high-resolution NeRF methods, our approach improves the rendering quality ofhigh-frequency details and achieves better visual effects in 4K high-resolution scenes.