cs.CVNov 3, 2026

Multi-Path Mamba Diffusion for 3D Point Cloud Semantic Scene Completion

Authors: Li Liang, Naveed Akhtar, Jordan Vice, Anwaar Ulhaq, Ajmal Saeed Mian

Organizations: The University of Western Australia · The University of Melbourne · Central Queensland University

Abstract

3D semantic scene completion plays a critical role in enabling various downstream tasks within autonomous systems by estimating both missing geometric structures and semantic information in scene data. In this work, we introduce a unique neural framework that integrates state space modeling with diffusion-based generative approaches to enhance 3D point cloud semantic scene completion. We propose MPath Mamba diffusion network, leveraging a skip-triple Mamba structure, dimensional decomposition residuals, and multi-directional dilations within multi-path data to efficiently process long-sequence data for feature extraction. The proposed method operates in the conditioned latent space of a variational autoencoder, enhanced by a state-space mechanism to improve 3D point cloud representation and reconstruction. Extensive experiments conducted on the SemanticKITTI benchmark demonstrate that our approach achieves highly competitive performance for 3D point cloud semantic scene completion.

Figures & tables

Explore similar work

CardsList
  1. PointDiffusion: Diffusion-Based Scene Completion in the Point Cloud Domain

    Jun 14, 2026Chidera Agbasiere, Mikhail Sannikov, Faith Ogunwoye +5Point Cloud Completion3D Reconstruction

  2. Deep Learning Approaches for 3D Medical Scene Completion: From Geometric Modeling to Generative Paradigms

    Jun 23, 2026Afifa Khaled, Said Jadid Abdulkadir, Majdy Mohamed Eltayeb Eltahir3D Gaussian Splatting3D Reconstruction

  3. ScaleBlind: Point Cloud Completion under Unknown Scale

    Sep 20, 2026Shenghui Wu, Chen Wang, Yuan Feng +3Point Cloud Completion