LatentAM：实时、大规模潜在高斯注意力映射通过在线字典学习

LatentAM is an online 3D Gaussian Splatting (3DGS) mapping framework designed to build scalable latent feature maps from streaming RGB-D observations for open-vocabulary robotic perception. It introduces an online dictionary learning approach that avoids the need for model-specific decoders to distill high-dimensional Vision-Language Model (VLM) embeddings, offering model-agnostic and pretraining-free advantages for plug-and-play integration with various VLMs at test time. Specifically, the method associates each Gaussian primitive with a compact query vector, which can be transformed into approximate VLM embeddings through an attention mechanism utilizing a learnable dictionary. The dictionary is efficiently initialized from streaming observations, ensuring real-time processing capabilities.