Loss Landscape Topology Reveals Why Simple Baselines are Competitive at 3D Point Cloud Segmentation Under Class Imbalance

2026-07-23

AI Summary

This study addresses the challenge of severe class imbalance in 3D point cloud semantic segmentation, where methods adapted from 2D vision often underperform for unclear reasons. The authors systematically evaluate eleven class-imbalance mitigation strategies across two mainstream 3D architectures and multiple datasets, complemented by error pattern diagnosis, decision boundary visualization, and geometric analysis of loss landscapes. They find that standard cross-entropy loss remains remarkably robust under both extreme (641:1) and moderate (56:1) imbalance, trailing specialized methods by only 0.8–3.3% mIoU. Crucially, this work reveals—through the lens of loss landscape topology—that class imbalance fundamentally reshapes the optimization geometry in 3D, explaining why direct transfers of 2D techniques fail and demonstrating that improper hyperparameter tuning can even degrade performance.

Abstract

Semantic segmentation of 3D point clouds faces severe class imbalance, yet the effectiveness of specialized imbalance-aware methods from 2D computer vision remains unclear in 3D contexts. We systematically evaluate 11 imbalance mitigation approaches across datasets with extreme (641:1) and moderate (56:1) imbalance ratios, revealing a surprising finding: standard cross-entropy with uniform weighting achieves competitive performance, typically within 0.8-3.3% mIoU of specialized methods across architectures and datasets. Through multifaceted mechanistic analysis of error patterns, decision boundaries, and the geometry of the optimization landscape, our analyses suggest that imbalance severity shapes the topology, creating narrow solution basins under extreme imbalance and flat plateaus under moderate imbalance. This appears to constrain the effectiveness of loss-level modifications, as all methods must navigate these geometric constraints. Our findings offer practical guidance; standard cross-entropy provides a robust baseline, with specialized methods offering modest improvements (0.8-3.3% mIoU) that vary by architecture and dataset but risk substantial degradation if poorly tuned. This work provides the first mechanistic explanation for why techniques proven effective in 2D do not readily transfer to point-based 3D point cloud segmentation, validated across two representative architectures.

Problem

• class imbalance
• 3D point cloud segmentation
• semantic segmentation
• loss landscape
• cross-domain generalization

Innovation

• loss landscape topology
• class imbalance
• 3D point cloud segmentation
• cross-entropy baseline
• geometric constraints