NGP-ERGAS: Revisit Instant Neural Graphics Primitives with the Relative Dimensionless Global Error in Synthesis

The newly emerging neural radiance fields (NeRF) methods can implicitly fulfill three-dimensional (3D) reconstruction via training a neural network to render novel-view images of a given scene with given posed images. The Instant Neural Graphics Primitives (Instant-NGP) method further improves the position encoding of NeRF. It obtains state-of-the-art efficiency. However, only a local pixel-wised loss is considered when training the Instant-NGP while overlooking the nonlocal structural information between pixels. Despite a good quantitative result, it leads to a poor visual effect, especially the completeness. Inspired by the stochastic structural similarity (S3IM) method that exploits nonlocal structural information of groups of pixels, this paper proposes a new method to improve the completeness of fast novel view synthesis. The proposed method first extends the thread-wised processing of the Instant-NGP to the processing in a custom thread block (i.e., a group of threads). Then, the relative dimensionless global error in synthesis, i.e., Erreur Relative Globale Adimensionnelle de Synthese (ERGAS), of a group of pixels corresponding to a group of threads is computed and incorporated into the loss function. Extensive experiments validate the proposed method. It can obtain better quantitative results than the original Instant-NGP with fewer iteration steps. PSNR is increased by 1%. Amazing qualitative results are obtained, especially for delicate structures and details such as lines and continuous structures. With the dramatic improvements in the visual effects, our method can boost the practicability of implicit 3D reconstruction in applications such as self-driving and augmented reality.