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.

Physics-guided computational imaging typically aggregates data fidelity, geometric reconstruction, and sensor consistency into a single scalar loss. In low signal-to-noise ratio (low-SNR) non-line-of-sight imaging, this centralized approach creates asymmetric gradient conflicts where the dominant constraints suppress physically meaningful updates. We propose treating multi-constraint training as a gradient coordination problem rather than scalar balancing. Our framework coordinates heterogeneous objectives through branch-wise gradient routing: soft conflict projection (PCGrad), hard physical constraint enforcement (PhysGuard), learnable sensor calibration, and a staged training protocol that decouples representation learning from nuisance parameter estimation. On held-out test scenes, the fully staged model improved the peak signal-to-noise ratio (PSNR) from 19.09 dB to 20.49 dB and the structural similarity index (SSIM) from 0.67 to 0.71 over the baseline, with consistent gains across the 48, 28, and 25 dB SNR levels. Qualitative evaluation on seven real-world scenes indicates sharper structure recovery and fewer artifacts. In this NLOS setting, gradient-level coordination is more reliable than scalar aggregation under heterogeneous constraints.