Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Point Cloud Semantic Segmentation with Sparse and Inhomogeneous Annotations

Zhiyi Pan, Nan Zhang, Wei Gao, Shan Liu, Ge Li

TL;DR

This work addresses the challenge of weakly supervised point cloud semantic segmentation when annotations are sparse and non-uniform. It introduces a gradient-sampling analysis with a probability density perspective and proposes Adaptive Annotation Distribution Network (AADNet) comprising Label-aware Downsampling (LaDS) and Multiplicative Dynamic Entropy with asynchronous training (MDE-AT) to mitigate gradient bias and epistemic uncertainty. The approach yields robust performance across S3DIS, ScanNetV2, and SemanticKITTI at very low label rates, demonstrating substantial improvements over existing methods in both uniform and non-uniform annotation scenarios. By enabling reliable segmentation under realistic labeling conditions, the method reduces annotation burden and improves generalization across diverse annotation distributions.

Abstract

Utilizing uniformly distributed sparse annotations, weakly supervised learning alleviates the heavy reliance on fine-grained annotations in point cloud semantic segmentation tasks. However, few works discuss the inhomogeneity of sparse annotations, albeit it is common in real-world scenarios. Therefore, this work introduces the probability density function into the gradient sampling approximation method to qualitatively analyze the impact of annotation sparsity and inhomogeneity under weakly supervised learning. Based on our analysis, we propose an Adaptive Annotation Distribution Network (AADNet) capable of robust learning on arbitrarily distributed sparse annotations. Specifically, we propose a label-aware point cloud downsampling strategy to increase the proportion of annotations involved in the training stage. Furthermore, we design the multiplicative dynamic entropy as the gradient calibration function to mitigate the gradient bias caused by non-uniformly distributed sparse annotations and explicitly reduce the epistemic uncertainty. Without any prior restrictions and additional information, our proposed method achieves comprehensive performance improvements at multiple label rates and different annotation distributions.

Point Cloud Semantic Segmentation with Sparse and Inhomogeneous Annotations

TL;DR

This work addresses the challenge of weakly supervised point cloud semantic segmentation when annotations are sparse and non-uniform. It introduces a gradient-sampling analysis with a probability density perspective and proposes Adaptive Annotation Distribution Network (AADNet) comprising Label-aware Downsampling (LaDS) and Multiplicative Dynamic Entropy with asynchronous training (MDE-AT) to mitigate gradient bias and epistemic uncertainty. The approach yields robust performance across S3DIS, ScanNetV2, and SemanticKITTI at very low label rates, demonstrating substantial improvements over existing methods in both uniform and non-uniform annotation scenarios. By enabling reliable segmentation under realistic labeling conditions, the method reduces annotation burden and improves generalization across diverse annotation distributions.

Abstract

Utilizing uniformly distributed sparse annotations, weakly supervised learning alleviates the heavy reliance on fine-grained annotations in point cloud semantic segmentation tasks. However, few works discuss the inhomogeneity of sparse annotations, albeit it is common in real-world scenarios. Therefore, this work introduces the probability density function into the gradient sampling approximation method to qualitatively analyze the impact of annotation sparsity and inhomogeneity under weakly supervised learning. Based on our analysis, we propose an Adaptive Annotation Distribution Network (AADNet) capable of robust learning on arbitrarily distributed sparse annotations. Specifically, we propose a label-aware point cloud downsampling strategy to increase the proportion of annotations involved in the training stage. Furthermore, we design the multiplicative dynamic entropy as the gradient calibration function to mitigate the gradient bias caused by non-uniformly distributed sparse annotations and explicitly reduce the epistemic uncertainty. Without any prior restrictions and additional information, our proposed method achieves comprehensive performance improvements at multiple label rates and different annotation distributions.
Paper Structure (14 sections, 7 equations, 8 figures, 5 tables)

This paper contains 14 sections, 7 equations, 8 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Preliminaries
  5. Gradient Sampling Approximation Analysis
  6. Label-aware Downsampling Strategy
  7. Multiplicative Dynamic Entropy
  8. Labeling Implementation
  9. Experiments
  10. Experiment Setting
  11. Comparison Results
  12. Ablation Study
  13. Further Analysis
  14. Conclusion

Figures (8)

  • Figure 1: Visual comparison about full annotations, uniform sparse annotations, and non-uniform sparse annotations. Based on the data presented in regarding the number of labeled categories per scene, the distribution disparity between non-uniform sparse annotation and full annotation is pronounced.
  • Figure 2: The framework of AADNet consists of two modules: (1) Label-aware downsampling strategy (LaDS) to boost the training annotation rate and maintain structural diversity. (2) Multiplicative dynamic entropy with asynchronous training (MDE-AT) to correct for gradient bias due to sparse labeling distributions as well as to explicitly improve epistemic certainty.
  • Figure 3: Compared with (a) voxel-level label-first downsampling, (b) voxel-level random downsampling retains diverse scene structural information.
  • Figure 4: Schematic diagram of gradient calibration by MDE.
  • Figure 5: The sparse annotations with different $G$.
  • ...and 3 more figures