LaRI: Layered Ray Intersections for Single-view
3D Geometric Reasoning

Rui Li1     Biao Zhang1     Zhenyu Li1     Federico Tombari2,3     Peter Wonka1    

ICML 2026

1KAUST     2Google     3Technical University of Munich    
Paper Code HuggingFace logo Demo Dataset

TL;DR: A single-feed-forward method that models unseen 3D geometry using layered point maps.

Descriptive alt text Descriptive alt text
Descriptive alt text Descriptive alt text
Descriptive alt text Descriptive alt text
Descriptive alt text Descriptive alt text

Summary

  • We model scene geometry using layered point maps, where each layer represents the intersection points through which rays pass on a surface.
  • We unify object- and scene-level geometric reasoning into a standard 2D regression task, achieving competitive or improved performance in a single feed-forward pass.
  • We provide a dataset construction, cleaning, and evaluation benchmark for this task.

Overview

Left: multiple ray-surface intersections + point maps ⇒ Unseen geometry representation.
Right: LaRI map \(\mathbf{V} \in \mathbb{R}^{H\times W \times L \times 3}\) + valid ray intersection mask \(\mathbf{M} \in \{0,1\}^{H\times W \times L}\) ⇒ Final 3D point cloud.

Layered Depth Visualization

Input Scene
Layered Depth Map

Layered Depth: 1

Object-level 3D Interactive Results

All models are subsampled to 20K points. Unseen geometry is higlighted with random color.

Scene-level 3D Interactive Results

All models are subsampled to 20K points. Unseen geometry is higlighted with random color.

Example 1
Example 2
Example 3
Example 4

3D Reconstruction & Reasoning Evaluation

In object-level comparison, LaRI yields comparable or better performances than existing large generative models. In scene-level comparison, LaRI achieves the best performance in unseen geometry estimation.

Computaion Efficiency

LaRI adopts a relaively lightweight solution with faster speed.

BibTeX

@inproceedings{li2026lari,
  title={LaRI: Layered Ray Intersections for Single-view 3D Geometric Reasoning},
  author={Li, Rui and Zhang, Biao and Li, Zhenyu and Tombari, Federico and Wonka, Peter},
  booktitle={Proceedings of the 43rd International Conference on Machine Learning},
  series={Proceedings of Machine Learning Research},
  publisher={PMLR},
  year={2026}
}

awesome webpage template