Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Topology-Guided Knowledge Distillation for Efficient Point Cloud Processing

Luu Tung Hai, Thinh D. Le, Zhicheng Ding, Qing Tian, Truong-Son Hy

TL;DR

This work tackles efficient point cloud processing by distilling a high-capacity teacher (Point Transformer V3) into a lightweight student through topology-aware representations and gradient-guided learning. It introduces a topology-based regularizer using Vietoris-Rips filtrations and persistence diagrams, combined with a gradient-driven feature alignment loss, and a distribution-level KL divergence term to guide knowledge transfer. The composite distillation objective preserves global geometric structure while emphasizing task-relevant local features, enabling a 16-fold reduction in parameters and ~$1.9\times$ faster inference with only modest accuracy loss on LiDAR semantic segmentation tasks across Nuscenes, SemanticKITTI, and Waymo. The approach achieves state-of-the-art KD performance on NuScenes and demonstrates strong generalization and practicality for edge deployments, with a public implementation available for reproducibility.

Abstract

Point cloud processing has gained significant attention due to its critical role in applications such as autonomous driving and 3D object recognition. However, deploying high-performance models like Point Transformer V3 in resource-constrained environments remains challenging due to their high computational and memory demands. This work introduces a novel distillation framework that leverages topology-aware representations and gradient-guided knowledge distillation to effectively transfer knowledge from a high-capacity teacher to a lightweight student model. Our approach captures the underlying geometric structures of point clouds while selectively guiding the student model's learning process through gradient-based feature alignment. Experimental results in the Nuscenes, SemanticKITTI, and Waymo datasets demonstrate that the proposed method achieves competitive performance, with an approximately 16x reduction in model size and a nearly 1.9x decrease in inference time compared to its teacher model. Notably, on NuScenes, our method achieves state-of-the-art performance among knowledge distillation techniques trained solely on LiDAR data, surpassing prior knowledge distillation baselines in segmentation performance. Our implementation is available publicly at: https://github.com/HySonLab/PointDistill

Topology-Guided Knowledge Distillation for Efficient Point Cloud Processing

TL;DR

This work tackles efficient point cloud processing by distilling a high-capacity teacher (Point Transformer V3) into a lightweight student through topology-aware representations and gradient-guided learning. It introduces a topology-based regularizer using Vietoris-Rips filtrations and persistence diagrams, combined with a gradient-driven feature alignment loss, and a distribution-level KL divergence term to guide knowledge transfer. The composite distillation objective preserves global geometric structure while emphasizing task-relevant local features, enabling a 16-fold reduction in parameters and ~ faster inference with only modest accuracy loss on LiDAR semantic segmentation tasks across Nuscenes, SemanticKITTI, and Waymo. The approach achieves state-of-the-art KD performance on NuScenes and demonstrates strong generalization and practicality for edge deployments, with a public implementation available for reproducibility.

Abstract

Point cloud processing has gained significant attention due to its critical role in applications such as autonomous driving and 3D object recognition. However, deploying high-performance models like Point Transformer V3 in resource-constrained environments remains challenging due to their high computational and memory demands. This work introduces a novel distillation framework that leverages topology-aware representations and gradient-guided knowledge distillation to effectively transfer knowledge from a high-capacity teacher to a lightweight student model. Our approach captures the underlying geometric structures of point clouds while selectively guiding the student model's learning process through gradient-based feature alignment. Experimental results in the Nuscenes, SemanticKITTI, and Waymo datasets demonstrate that the proposed method achieves competitive performance, with an approximately 16x reduction in model size and a nearly 1.9x decrease in inference time compared to its teacher model. Notably, on NuScenes, our method achieves state-of-the-art performance among knowledge distillation techniques trained solely on LiDAR data, surpassing prior knowledge distillation baselines in segmentation performance. Our implementation is available publicly at: https://github.com/HySonLab/PointDistill
Paper Structure (26 sections, 1 theorem, 10 equations, 4 figures, 7 tables)

This paper contains 26 sections, 1 theorem, 10 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. 3D Point Cloud Processing
  4. Point Transformer Architecture
  5. Knowledge Distillation
  6. Methodology
  7. Overview of the Framework
  8. Topology-Aware Distillation Learning
  9. Gradient-Guided Knowledge Distillation
  10. Overall Distillation Objective
  11. Experiments and Results
  12. Experimental Results
  13. Comprehensive Teacher - Student Analysis
  14. Discussion and Future Work
  15. Experimental Details
  16. ...and 11 more sections

Key Result

Theorem 1

Let $D_T$ and $D_S$ be the persistence diagrams for the teacher and student models, respectively. If $\mathcal{L}_{\text{CD}}(D_T, D_S) \to 0$, then the 2-Wasserstein distance between $D_T$ and $D_S$ also converges: Thus, minimizing Chamfer Distance implicitly minimizes the Wasserstein distance between persistence diagrams, ensuring topological consistency.

Figures (4)

  • Figure 1: Comprehensive comparisons between our proposed method and state-of-the-art knowledge distillation baselines across multiple evaluation metrics point-to-voxeltpv-igkdumkdcmdfusion. (a) The radar chart demonstrates that our method achieves consistently better mIoU on three key datasets (NuScenes, SemanticKITTI, and Waymo), along with favorable FPS and memory efficiency. (b)-(d) As the input point token length increases, our approach maintains lower GPU memory usage and FLOPs, while sustaining significantly faster inference speed. (c) Peak GPU memory usage during inference, measured using max memory allocated function in Torch. This metric reflects the highest amount of memory used by the PyTorch tensors by the caching allocator during the inference phase. Notably, this value may differ significantly from the memory reported by the PyTorch Profiler due to its inclusion of temporary allocations used by CUDA kernels.
  • Figure 2: Overview of the proposed knowledge distillation framework for point cloud processing. The framework transfers knowledge from a teacher model (yellow) to a student model (green) using topological data analysis (TDA) (1), KLD matching (2), student semantic segmentation loss (3), and gradient-guided feature alignment (4). The total loss is the combination of (1), (2), (3) and gradient-guided feature loss.
  • Figure 3: Illustration of topology-aware analysis through Vietoris–Rips filtration. The top row depicts the evolution of the simplicial complex as the filtration parameter $\epsilon$ increases. The bottom part shows the corresponding barcode representation of persistent homology groups in different dimensions ($H_0, H_1, H_2$).
  • Figure 4: Visualization of our method on the nuScenes validation set. (a) Ground truth, (b) teacher model prediction, and (c) student model prediction. The student model closely follows the teacher’s output and ground truth, successfully capturing almost all object classes, demonstrating the effectiveness of the knowledge distillation process.

Theorems & Definitions (2)

  • Theorem 1
  • proof