Diff2DGS: Reliable Reconstruction of Occluded Surgical Scenes via 2D Gaussian Splatting

Diff2DGS is aimed at a problem that matters anywhere real-time reconstruction has to survive heavy occlusion: surgical video. The pipeline first uses a diffusion-based video module with temporal priors to fill in tissue hidden by instruments, then feeds that into an adapted 2D Gaussian Splatting setup built for dynamic deformation and anatomy.

The graphics angle is the interesting part for developers: 2DGS is being pushed beyond pretty reconstructions and into geometry that can hold up under motion, occlusion, and changing tissue shape. A Learnable Deformation Model helps track dynamic surfaces, while an adaptive depth weight is used to improve geometric fidelity instead of optimizing for image quality alone.

That distinction matters because the work also adds depth evaluation on the SCARED dataset, extending beyond the usual image metrics used by EndoNeRF and StereoMIS-style benchmarks. The results suggest that a reconstruction can look good while still being wrong in 3D, which is a useful warning for anyone building neural rendering or scene reconstruction systems.

For game developers, the broader takeaway is that the same class of techniques used for real-time scene capture, avatar reconstruction, or mixed-reality content can benefit from explicit handling of occlusion and depth consistency. The code is available, and the paper positions the method as a step toward more reliable reconstruction in visually messy, dynamic environments.