Go to ICLR 2026 Conference homepageIPG-Rec: Instance-aware Progressive Geometry Rectification for High-fidelity Sparse Novel View Synthesis
Keywords: gaussian splatting, instance-aware depth supervision, geometry rectification, sparse novel view synthesis
Abstract: Sparse novel view synthesis is a challenging problem due to the limited information available. While recent 3D Gaussian Splatting (3DGS) approaches have leveraged monocular depth or diffusion priors to improve reconstruction quality, they struggle to generate multi-view consistent geometry efficiently. To address this problem, we propose an Instance-aware Progressive Geometry Rectification method, namely PG-Rec, to reconstruct the high-fidelity geometry from sparse inputs. Notably, our approach progressively and jointly optimizes 3D Gaussian representations by leveraging reliable pseudo-view images, along with instance-level and scene-level depth regularization, which promotes the reconstruction of high-fidelity 3D geometry with implicit cross-view semantic consistency. Considering insufficient information from sparse views, we employ instance-level and scene-level depth regularization to refine the 3D geometry cooperatively. The instance depth guides the 3D Gaussians to move toward their corresponding object, while the global depth maintains the relative spatial positions of Gaussians in different instances. With geometry refined by depth regularization, 3DGS renders more realistic images that guide diffusion to generate reliable pseudo-views. These pseudo-views are then used to further refine geometry. By combining depth regularization with high-fidelity pseudo-view rendering, our method progressively mitigates reconstruction defects from sparse inputs and acquires high-fidelity rendering images. Extensive experiments demonstrate that our PIGR outperforms current state-of-the-art methods in sparse novel view synthesis.
Primary Area: applications to computer vision, audio, language, and other modalities
Submission Number: 6864
Loading