I2PRef: Image-Driven Point Completion with Iterative Refinement

Researchers introduced I2PRef, a single-image point cloud completion method for 3D reconstruction.

The pipeline first predicts a coarse point cloud from one RGB image, then refines it with transformer-based image and point features.

On ShapeNet-ViPC, it achieved state-of-the-art results and a 12.3% relative gain in Chamfer Distance.

The work shows that high-quality point cloud completion is possible without any 3D input, advancing image-based 3D understanding.