DCRR++: Unsupervised Reflection Removal and Novel View Synthesis via Dual-Pixel Guided 3D Gaussian Splatting

Image de-reflection is a fundamental and critical task in computer vision. Existing monocular methods often struggle to separate the transmission and reflection layers due to a lack of depth cues, particularly under strong illumination or multi-layer reflection scenarios. Although recent advances, such as 3D Gaussian Splatting (3DGS), have shown potential in layer separation via novel view synthesis, they remain limited by the inherent ambiguity of monocular inputs and fail to accurately capture viewpoint-dependent reflection components in unconstrained environments. In this work, we simplify the de-reflection task by combining Dual-Pixel (DP) technology with 3DGS, forming the first unsupervised framework to unlock the potential of spatial perception and layer separation. Specifically, our Dual-View Coordinated Reflection Removal (DCRR++) framework integrates valuable depth cues from DP sensors with 3DGS rendering, transforming the complex geometric-photometric decoupling process into a simple 3D spatial Gaussian pruning problem. DCRR++ first estimates the transmission layer and opacity via differentiable rasterization. To model dynamic reflection, we introduce a Reflection-Aware Charging Module (RACM) with a synergistic dual-branch (geometric and texture) architecture that reconstructs anisotropic reflection layers through pose embedding and feature injection. This collaborative approach translates the challenging reflection generation problem from scratch to a more efficient reflection residual learning task. Further, we propose Dual-Pixel-Driven Reflection Gaussian Pruning (DPRGP) to refine the separation process. By leveraging the physical properties of DP sensors, our method identifies differences in 3D distributions between the dual views to generate potential reflection candidate Gaussian primitives with significant disparity, which are then progressively pruned from the scene to achieve reflection removal in unconstrained scenarios. Two real-world DP-based datasets that include paired reflection/reflection-free images have been collected. Extensive experiments demonstrate our DCRR++, as an unsupervised model, has competitive performance compared to state-of-the-art methods, even outperforming most supervised monocular single-image reflection removal methods on certain datasets.