Learnable Semantic Water-Filling for Task-Oriented 3D Point Cloud Transmission

Task-oriented point cloud transmission preserves only the critical semantic information required by downstream inference tasks rather than reconstructing entire geometric point sets. However, existing deep joint source-channel coding (JSCC) frameworks typically allocate wireless communication resources uniformly across latent features, discarding spatial semantic variations. This letter presents Learnable Semantic Water-Filling (LSWF), a mathematically principled, differentiable feature allocation optimization paradigm. LSWF extracts point-wise semantic importance from latent spaces via an ultra-lightweight predictor. A learnable water-filling operator then solves a closed-form rate-distortion allocation problem dynamically, concentrating transmission energy on task-critical point geometries while naturally generating highly sparse structural representations. The entire framework is fully differentiable, permitting joint end-to-end backpropagation with deep neural backbones and channel-state-aware layers. Evaluated on the ShapeNet part segmentation benchmark, LSWF achieves state-of-the-art accuracy under both AWGN and Rayleigh fading channels while compressing the effective transmission rate by over 97%. This framework offers an efficient signal processing solution for non-uniform resource allocation in intelligent edge systems.