The paper introduces a Bilateral Propagation Network (BP-Net) for depth completion, aiming to derive a dense depth map from sparse depth measurements and a synchronized color image. Unlike existing methods that primarily focus on iterative refinement of initial depth estimates, BP-Net propagates depth information from nearby measurements in the early stage, avoiding direct convolution on sparse data. The key innovation is a non-linear propagation model where the output depth is a weighted combination of nearby valid depth measurements, with coefficients generated by a multi-layer perceptron (MLP) conditioned on both radiometric difference and spatial distance. This approach ensures that depth propagation prefers the nearest values in both domain and range, enhancing the effectiveness of the multi-modal fusion and depth refinement stages. The method is evaluated on both indoor (NYUv2) and outdoor (KITTI) datasets, achieving state-of-the-art performance and demonstrating the importance of early-stage propagation. The code and trained models are available online.The paper introduces a Bilateral Propagation Network (BP-Net) for depth completion, aiming to derive a dense depth map from sparse depth measurements and a synchronized color image. Unlike existing methods that primarily focus on iterative refinement of initial depth estimates, BP-Net propagates depth information from nearby measurements in the early stage, avoiding direct convolution on sparse data. The key innovation is a non-linear propagation model where the output depth is a weighted combination of nearby valid depth measurements, with coefficients generated by a multi-layer perceptron (MLP) conditioned on both radiometric difference and spatial distance. This approach ensures that depth propagation prefers the nearest values in both domain and range, enhancing the effectiveness of the multi-modal fusion and depth refinement stages. The method is evaluated on both indoor (NYUv2) and outdoor (KITTI) datasets, achieving state-of-the-art performance and demonstrating the importance of early-stage propagation. The code and trained models are available online.