Abstract: To address the issue of excessive parameter volume in RGB-D pose estimation feature extraction networks, the lightweight subnetwork—light weight net (LWNet) is proposed to replace the ResNet18 backbone in XYZNet. First, an improved C2f module was designed in conjunction with the partial convolution (PConv) module to form a feature extraction structure for shallow-level feature extraction, enhancing feature reusability.Second, the feature fusion efficient layer aggregation networks (FF-ELAN) module was designed for deep-level feature extraction, integrating global features in a lightweight manner.Afterward, three downsampling structures distinct from ResNet18 were designed, cutting the number of parameters while boosting performance. Experimental results show that the optimized network achieves a 58.1% reduction in floating-point operations (FLOPs), a 57.3% decrease in parameter count, and a 10.2% improvement in inference speed. Simultaneously, it achieves accuracy improvements of 1.3 percentage points on the LineMOD dataset and 0.8 percentage points on the YCB-Video dataset. This architecture significantly enhances deployment potential for edge computing scenarios.