NeuSE: Neural SE(3)-Equivariant Embedding for Long-Term Object-Based Simultaneous Localization and Mapping

NeuSE generates SE(3) and scale-equivariant latent embeddings from partial object observations to serve as compact point cloud surrogates. These embeddings encode shape, scale, and transform information, allowing relative frame transforms to be derived directly from latent codes. This enables consistent object-based SLAM in environments with varying object sizes and poses.