NTT Paper Accepted for ICPR 2026, a Premier International Conference in Pattern Recognition

NTT's paper was accepted at ICPR 2026, the International Conference on Challenges in the Fields of Pattern Recognition, Computer Vision and Machine Learning, to be held in Lyon, France from August 17, 2026 to 22nd. The adopted papers are as follows.

Abbreviated names of the laboratories:
HI: NTT Human Informatics Laboratories

■Training-Free Photo-Realistic Point Cloud Rendering via Geometry-Aware Densification and Multi-View Refinement

Shogo Sato (HI), Kazuhiko Murasaki (HI), Ryuichi Tanida (HI)

We propose a training-free and low-cost rendering method for generating photo-realistic images from colored point clouds. Colored point clouds captured by LiDAR are sparse, so they are not suitable for producing high-quality images directly. In our method, we first upsample the point cloud and geometrically assign 3D Gaussian Splatting parameters to each point. This enables dense and natural image rendering. We also use pre-trained models to restore missing or blurry regions. In addition, multi-view refinement helps recover structures that are difficult to reconstruct from a single view while maintaining geometric consistency. Experiments show that our method achieves rendering quality comparable to conventional learning-based methods without requiring training data and with lower computational cost.