Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EAGLE: Contextual Point Cloud Generation via Adaptive Continuous Normalizing Flow with Self-Attention

Linhao Wang, Qichang Zhang, Yifan Yang, Hao Wang, Ye Su

TL;DR

This work proposes EAGLE, a probabilistic generative model that integrates self-attention mechanisms with adaptive continuous normalizing flows, and introduces an adaptive bias correction mechanism within flow-based models, which dynamically adjusts to different input contexts and alleviates bias-drift issues arising from standard initialization.

Abstract

As 3D point clouds become the prevailing shape representation in computer vision, generating high-quality point clouds remains a challenging problem. Flow-based models have shown strong potential due to exact likelihood estimation and invertible mappings. However, existing flow-based methods for point clouds typically rely on point-wise feature extractors, which limits their ability to model long-range dependencies and global structural relationships among points. Inspired by the wide adoption of Transformers, we explored the complementary roles of self-attention mechanisms, CNN and flow-based model. To this end, we propose EAGLE, a probabilistic generative model that integrates self-attention mechanisms with adaptive continuous normalizing flows. The self-attention module explicitly models pairwise dependencies among points, enabling effective capture of global contextual information. In addition, we introduce an adaptive bias correction mechanism within flow-based models, which dynamically adjusts to different input contexts and alleviates bias-drift issues arising from standard initialization. Extensive experiments on ShapeNet and ModelNet datasets demonstrate the effectiveness of our proposed method.

EAGLE: Contextual Point Cloud Generation via Adaptive Continuous Normalizing Flow with Self-Attention

TL;DR

This work proposes EAGLE, a probabilistic generative model that integrates self-attention mechanisms with adaptive continuous normalizing flows, and introduces an adaptive bias correction mechanism within flow-based models, which dynamically adjusts to different input contexts and alleviates bias-drift issues arising from standard initialization.

Abstract

As 3D point clouds become the prevailing shape representation in computer vision, generating high-quality point clouds remains a challenging problem. Flow-based models have shown strong potential due to exact likelihood estimation and invertible mappings. However, existing flow-based methods for point clouds typically rely on point-wise feature extractors, which limits their ability to model long-range dependencies and global structural relationships among points. Inspired by the wide adoption of Transformers, we explored the complementary roles of self-attention mechanisms, CNN and flow-based model. To this end, we propose EAGLE, a probabilistic generative model that integrates self-attention mechanisms with adaptive continuous normalizing flows. The self-attention module explicitly models pairwise dependencies among points, enabling effective capture of global contextual information. In addition, we introduce an adaptive bias correction mechanism within flow-based models, which dynamically adjusts to different input contexts and alleviates bias-drift issues arising from standard initialization. Extensive experiments on ShapeNet and ModelNet datasets demonstrate the effectiveness of our proposed method.

Paper Structure

This paper contains 17 sections, 22 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Model Background
  4. Overview of Flow-Based Models
  5. Continuous Normalizing Flow
  6. Variational Autoencoder
  7. Fundamental Training Framework
  8. Method
  9. Context-Aware Encoder with Self-Attention
  10. Adaptive Continuous Normalizing Flow
  11. Training Objective
  12. Experiments
  13. Experimental Setup
  14. Point Cloud Generation
  15. Representation Learning Evaluation
  16. ...and 2 more sections

Figures (4)

  • Figure 1: The visualization of the proposed method. (A) illustrates how the objective is computed during the training stage. (B) illustrates the sampling (generation) stage. The ConvBlock consists of 1D convolution layers, batchnorm layers and ReLU activation function. The latent block maps global point cloud features to the parameters of a latent distribution, from which latent variables are sampled via the reparameterization trick for subsequent generative modeling.
  • Figure 2: Visualization of our generation results (EAGLE) compared to baselines. EAGLE generates comparable and high-quality point clouds.
  • Figure 3: Visualization of the structure of our proposed Self-Attention block.
  • Figure 4: More visualization results of point clouds generated by our model. From top to bottom: airplane, car, chair.