6D object pose estimation is crucial for applications such as robotic manipulation and augmented reality, yet it remains highly challenging when dealing with objects of significantly different aspect ratios or the drastic appearance variations of a single object caused by pose changes. Most existing methods focus on designing more complex backend fusion modules, while largely overlooking a fundamental problem at the feature extraction frontend: the geometric mismatch between the fixed, square receptive fields of standard convolutions and the varied projected morphologies of objects. This mismatch, along with noise in fused features and ambiguity in regression, limits the performance ceiling of current methods. To this end, this paper proposes a novel Pose-Perceptive Convolution (PPC) and constructs a new Pose-Perceptive Fusion Network (PPF-Net). Its core component, the Pose-Perceptive Convolution, fundamentally resolves the aforementioned geometric mismatch by dynamically adapting the shape and sampling density of its receptive field. Experiments on four benchmarks show that PPF-Net improves the VSD score by 19.4% over FFB6D on MP6D, and achieves 96.7% ADD-S on YCB-Video, approaching state-of-the-art accuracy. Crucially, these gains are realized with minimal computational overhead, avoiding the heavy latency of backend-intensive approaches. This validates that frontend feature extraction is an efficient strategy for robust 6D pose estimation.

6D object pose estimation from RGB-D data is a core problem in robotic perception. Representative approaches such as DenseFusion fuse per-pixel RGB and depth features, yet they still face challenges under heavy occlusion and in cluttered backgrounds. This paper revisits the DenseFusion architecture from the perspective of feature enhancement and global context modeling and proposes an enhanced network named RAFusion, which strengthens feature representation through two complementary modules. First, Squeeze-and-Excitation (SE) blocks are integrated into the RGB feature extractor to adaptively reweight channel responses, enhancing salient appearance cues while suppressing noise. Second, a RealFormer-based block is introduced before global pooling, applying residual self-attention to the fused RGB-D features for more effective global context modeling. Experiments on the LINEMOD and Occlusion LINEMOD datasets show that RAFusion improves the DenseFusion baseline and achieves competitive performance compared with representative 6D pose estimation methods. Ablation studies and computational analysis further support the effectiveness of the proposed modules.