SRFlow3DNet: Monocular Facial Scene Flow with Optical Flow Branch Regularization

Existing scene flow datasets are mainly designed for general or autonomous driving scenarios, lacking high-resolution, geometrically consistent facial motion data, which limits effective modeling of 3D facial motion. In this research, a monocular RGB (Red, Green and Blue) facial scene flow estimation model is studied , aiming to accurately predict per-pixel 3D motion in facial regions from consecutive image sequences. To this end, the SRFlow3D dataset (Splatting Rasterization Flow 3D) is proposed, which leverages dynamic 3D Gaussian representations with splatting-based rasterization to simultaneously render RGB images, optical flow, and depth-related supervision data. Based on the SRFlow3D dataset, the SRFlow3DNet (Splatting Rasterization Guided Flow 3D Network) is proposed, which jointly predicts optical flow and depth variation along the viewing direction under monocular RGB input to obtain per-pixel 3D scene flow, and introduces optical flow branch regularization to enhance geometric and temporal consistency of non-rigid facial motion. Experimental results show that, compared with existing scene flow estimation methods, SRFlow3DNet reduces the optical flow metric End-Point Error (EPE) from 0.4984 to 0.3768 and the scene flow metric 3D End-Point Error (EPE3D) from 1.0826 to 0.4308, achieving significant performance gains in the monocular RGB facial scene flow estimation task.