Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Loss Distillation via Gradient Matching for Point Cloud Completion with Weighted Chamfer Distance

Fangzhou Lin, Haotian Liu, Haoying Zhou, Songlin Hou, Kazunori D Yamada, Gregory S. Fischer, Yanhua Li, Haichong K. Zhang, Ziming Zhang

TL;DR

It is observed that with proper weighted functions, the weighted CD can always achieve similar performance to HyperCD, and the Landau weighted CD, namely Landau CD, can outperform HyperCD for point cloud completion and lead to new state-of-the-art results on several benchmark datasets.

Abstract

3D point clouds enhanced the robot's ability to perceive the geometrical information of the environments, making it possible for many downstream tasks such as grasp pose detection and scene understanding. The performance of these tasks, though, heavily relies on the quality of data input, as incomplete can lead to poor results and failure cases. Recent training loss functions designed for deep learning-based point cloud completion, such as Chamfer distance (CD) and its variants (\eg HyperCD ), imply a good gradient weighting scheme can significantly boost performance. However, these CD-based loss functions usually require data-related parameter tuning, which can be time-consuming for data-extensive tasks. To address this issue, we aim to find a family of weighted training losses ({\em weighted CD}) that requires no parameter tuning. To this end, we propose a search scheme, {\em Loss Distillation via Gradient Matching}, to find good candidate loss functions by mimicking the learning behavior in backpropagation between HyperCD and weighted CD. Once this is done, we propose a novel bilevel optimization formula to train the backbone network based on the weighted CD loss. We observe that: (1) with proper weighted functions, the weighted CD can always achieve similar performance to HyperCD, and (2) the Landau weighted CD, namely {\em Landau CD}, can outperform HyperCD for point cloud completion and lead to new state-of-the-art results on several benchmark datasets. {\it Our demo code is available at \url{https://github.com/Zhang-VISLab/IROS2024-LossDistillationWeightedCD}.}

Loss Distillation via Gradient Matching for Point Cloud Completion with Weighted Chamfer Distance

TL;DR

It is observed that with proper weighted functions, the weighted CD can always achieve similar performance to HyperCD, and the Landau weighted CD, namely Landau CD, can outperform HyperCD for point cloud completion and lead to new state-of-the-art results on several benchmark datasets.

Abstract

3D point clouds enhanced the robot's ability to perceive the geometrical information of the environments, making it possible for many downstream tasks such as grasp pose detection and scene understanding. The performance of these tasks, though, heavily relies on the quality of data input, as incomplete can lead to poor results and failure cases. Recent training loss functions designed for deep learning-based point cloud completion, such as Chamfer distance (CD) and its variants (\eg HyperCD ), imply a good gradient weighting scheme can significantly boost performance. However, these CD-based loss functions usually require data-related parameter tuning, which can be time-consuming for data-extensive tasks. To address this issue, we aim to find a family of weighted training losses ({\em weighted CD}) that requires no parameter tuning. To this end, we propose a search scheme, {\em Loss Distillation via Gradient Matching}, to find good candidate loss functions by mimicking the learning behavior in backpropagation between HyperCD and weighted CD. Once this is done, we propose a novel bilevel optimization formula to train the backbone network based on the weighted CD loss. We observe that: (1) with proper weighted functions, the weighted CD can always achieve similar performance to HyperCD, and (2) the Landau weighted CD, namely {\em Landau CD}, can outperform HyperCD for point cloud completion and lead to new state-of-the-art results on several benchmark datasets. {\it Our demo code is available at \url{https://github.com/Zhang-VISLab/IROS2024-LossDistillationWeightedCD}.}
Paper Structure (26 sections, 7 equations, 7 figures, 7 tables, 2 algorithms)

This paper contains 26 sections, 7 equations, 7 figures, 7 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Training Loss
  3. Motivation
  4. Approach: Loss Distillation via Gradient Matching
  5. Contributions
  6. Related Work
  7. Distance Metrics for Point Cloud Completion
  8. Knowledge Distillation (KD)
  9. Weighted Chamfer Distance
  10. Our Approach
  11. Chamfer Distance (CD)
  12. Hyperbolic Chamfer Distance (HyperCD)
  13. Loss Distillation with Weighted CD
  14. Optimization
  15. Bilevel Optimization with Weighted CD
  16. ...and 11 more sections

Figures (7)

  • Figure 1: Illustration of distributions (with scaling and proper hyperparameters) that are similar to the gradient weighting distribution from HyperCD in backpropagation and can be taken as candidate weighting functions in weighted CD.
  • Figure 2: Illustration of (a) reference distance distribution from HyperCD, and (b-c) curve fitting using different approximations of $z^{(W)}$.
  • Figure 3: Visualization of the real-world(KITTI) benchmark (Row 1: sparse input, Row 2: HyperCD, Row 3: LandauCD).
  • Figure 4: Visualization of ShapeNet-55 benchmark. Gray represents the partial input. Yellow represents HyperCD. Green represents Landau CD.
  • Figure 5: Visual comparison on Shapenet-55. Row-1: Seedformer with HyperCD. Row-2: Seedformer with Landau CD.
  • ...and 2 more figures