NeurNCD: Novel Class Discovery via Implicit Neural Representation
Junming Wang, Yi Shi · 2024
Figure 1: NeurNCD leverages implicit neural representation, replacing traditional explicit 3D segmentation maps [19], to enhance the accuracy of novel class discovery.Specifically, the meticulously designed Embedding-NeRF model employs KL divergence, achieving the transfer and association of 2D-3D features while producing semantic embedding and entropy by aggregating information from multiple views.Then by integrating with other key components, i.e., feature query, feature modulation and clustering, to ultimately reconstruct refined, low-noise, and hole-free images and 3D structures.