Exploring dynamic plane representations for neural scene reconstruction

The efficient tri-plane representations present limited expressivity for encoding complex 3D scenes. To cope with the hampered spatial expressivity of tri-planes, this paper proposes a novel dynamic plane representation method for 3D scene reconstruction, including dynamic long-axis plane learning, a point-to-plane relationship module, and explicit coarse-to-fine feature projection. First, the proposed dynamic long-axis plane learning employs several planes along the principal axis and adapts planar positions dynamically, which can enhance geometry expressivity. Second, a point-to-plane relationship module is proposed to capture distinguished point features by learning the feature bias between plane features and point features. Third, the explicit coarse-to-fine feature projection employs a non-linear transformation to capture fine features from learnable coarse features, exploiting both local and global information with fewer increases in parameters. Experimental results on ScanNet and 7-Scenes demonstrate that our method achieves state-of-the-art performance with comparable computational costs.