Title: I2PRef: Image-Driven Point Completion with Iterative Refinement

Abstract: This study introduces an image-conditioned framework for point cloud completion, positioning the image as the fundamental source of geometric information rather than a supplementary cue. Central to our approach is the Image-to-Point (I2P) module, which is capable of generating complete point clouds directly from a single RGB image, eliminating the requirement for any 3D input data. Furthermore, we propose a transformer-based Point-to-Point (P2P) refinement stage that employs self- and cross-attention mechanisms between point tokens and image features to progressively enhance the initial coarse output from the I2P module. This architecture allows the image encoder to acquire robust geometric representations, while the P2P component systematically restores intricate details. In contrast to contemporary multimodal techniques that depend on fusion modules or auxiliary losses, our explicit I2P formulation establishes a potent, geometry-centric prior derived exclusively from visual data. Comprehensive evaluations on the ShapeNet-ViPC dataset reveal that our method achieves state-of-the-art completion results, delivering a 12.3% relative improvement in Chamfer Distance compared to previous approaches. The implementation code is accessible at: https://github.com/AzharSindhi/I2PRef.git

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC