SurfelFlow: Surface-Aware 2D Gaussian Streaming for Monocular Dynamic 4D Reconstruction

Recovering time-varying 3D scenes from monocular dynamic videos is challenging because each frame provides only a single view of a changing scene. Fast, large-magnitude motion further weakens geometric stability and image fidelity under monocular observations, especially when depth, motion, and visibility must be inferred from imperfect frame-wise priors. This paper proposes SurfelFlow, a surface-aware 2D Gaussian streaming 4D reconstruction method for this setting. SurfelFlow reduces the downstream geometric amplification of residual monocular-depth and optical-flow errors by changing how propagated primitives represent local surface support, rather than by replacing the external priors themselves. It represents local rendering primitives as 2D Gaussian surface patches aligned with the currently visible surface while retaining depth- and flow-guided frame-wise point propagation. It also introduces single-frame surface geometric regularization and streaming Gaussian visibility management to stabilize local surfaces and suppress residual historical dynamic Gaussians in disoccluded regions. Experiments and ablation studies on the public real-world DAVIS2017-dev dataset show that SurfelFlow improves the per-sequence macro-average PSNR over the GFlow baseline by 4.16 dB and outperforms the evaluated representative baselines in image reconstruction quality. Furthermore, evaluation against held-out registered sensor depth on four dynamic Bonn RGB-D sequences shows lower visible-surface errors, reducing scale-aligned absolute relative depth error (AbsRel) from 0.1241 to 0.1129 and surface normal mean angular error (MAE) from 58.55∘ to 54.51∘.