[Paper Review] AdvectiveNet: An Eulerian-Lagrangian Fluidic reservoir for Point Cloud Processing

Xingzhe He, Helen Lu Cao+1 more|arXiv (Cornell University)|Jan 31, 2020

Computer Graphics and Visualization Techniques51 references4 citations

TL;DR

AdvectiveNet introduces a novel Eulerian-Lagrangian fluidic reservoir for point cloud processing, modeling feature evolution via learned advection in a velocity field inspired by fluid dynamics. It achieves state-of-the-art performance on ModelNet40 and ShapeNet with significantly fewer parameters than competing methods.

ABSTRACT

This paper presents a novel physics-inspired deep learning approach for point cloud processing motivated by the natural flow phenomena in fluid mechanics. Our learning architecture jointly defines data in an Eulerian world space, using a static background grid, and a Lagrangian material space, using moving particles. By introducing this Eulerian-Lagrangian representation, we are able to naturally evolve and accumulate particle features using flow velocities generated from a generalized, high-dimensional force field. We demonstrate the efficacy of this system by solving various point cloud classification and segmentation problems with state-of-the-art performance. The entire geometric reservoir and data flow mimics the pipeline of the classic PIC/FLIP scheme in modeling natural flow, bridging the disciplines of geometric machine learning and physical simulation.

Motivation & Objective

• To address the challenges of constructing effective convolution stencils and evolving nonlinear features in point cloud learning.
• To bridge geometric machine learning and physical simulation by modeling point cloud processing as a fluid advection problem.
• To develop a hybrid Eulerian-Lagrangian framework that leverages mature computational fluid dynamics infrastructure for deep learning.
• To enable efficient, low-parameter learning by embedding feature evolution within a physically grounded reservoir.
• To demonstrate that physics-inspired reservoirs can match or exceed SOTA performance in point cloud classification and segmentation.

Proposed method

• The method employs a static Eulerian grid to represent world-space spatial structure and Lagrangian particles to track evolving features.
• Feature evolution is governed by advection in a learned velocity field, emulating fluid convection and enabling nonlinear feature separation.
• Particle-grid transfer is performed using interpolation schemes inspired by the PIC/FLIP methods in computational fluid dynamics.
• The system uses a generalized, high-dimensional force field to generate velocity fields that guide particle movement and feature accumulation.
• The entire network operates as a physical reservoir, with training equivalent to finding optimal control forces that minimize a loss energy function.
• The architecture supports global fast convolutions via the Eulerian grid and tracks non-linear feature evolution via Lagrangian particles.

Experimental results

Research questions

• RQ1 Can a physics-inspired fluidic reservoir with Eulerian-Lagrangian duality improve point cloud feature learning?
• RQ2 How can fluid dynamics principles like advection and particle tracking be adapted to enhance feature representation in point cloud networks?
• RQ3 Can a hybrid grid-particle system outperform existing point-based or grid-based networks with fewer parameters?
• RQ4 To what extent can the PIC/FLIP numerical framework be repurposed for deep learning in geometric data?
• RQ5 Does learning via physical advection dynamics lead to better generalization and performance in point cloud tasks?

Key findings

• On ModelNet10, AdvectiveNet achieved 95.4% accuracy, surpassing state-of-the-art methods using grid-based representations.
• On ModelNet40, it reached 92.8% accuracy, matching DGCNN (92.9%) while using only 1M parameters compared to DGCNN’s 21M.
• For ShapeNet part segmentation, AdvectiveNet achieved 86.2% mIoU, outperforming all point-based baselines with 1.1M parameters and only 2 time steps.
• The method demonstrated superior parameter efficiency, using 4% to 25% fewer parameters than SOTA models like PointCNN and DGCNN.
• Visualizations of the segmentation process showed coherent, physically plausible feature evolution over time.
• The framework successfully mimics the PIC/FLIP fluid simulation pipeline, enabling stable and effective feature propagation in high-dimensional space.