Image-Inpainting Guided Point Cloud Completion

Abstract Single object point cloud completion aims to recover complete object geometry from incomplete observations. Existing methods either rely solely on point cloud geometry or exploit auxiliary visual information, often re-quiring calibrated RGB images or other additional inputs. However, incomplete point clouds inherently lack fine-grained geometric and semantic cues, limiting reconstruction quality under severe missing structures. To address this issue, we propose a tri-modal framework that enriches point cloud completion with complementary depth and text priors. Specifically, we synthesize depth maps from partial point clouds through an image completion module and combine them with lightweight category-level text prompts to provide complementary structural and semantic guidance, without requiring calibrated RGB images. Furthermore, we construct a Point-Depth-Text triplet corpus based on existing point cloud completion datasets, enabling multimodal learning without additional manual annotation. Extensive experiments on the PCN, ShapeNet-55, and MVP benchmarks demonstrate that the proposed method achieves competitive results in both quantitative and qualitative evaluations.