How Sampling Strategy Affects Imbalance Mitigation in LiDAR Segmentation: A Study of Structured vs. Random Point-Based Architectures

Antonis Savva; Christos Kyrkou; Theocharis Theocharides

cs.CVAug 17, 2026

How Sampling Strategy Affects Imbalance Mitigation in LiDAR Segmentation: A Study of Structured vs. Random Point-Based Architectures

Authors: Antonis Savva, Christos Kyrkou, Theocharis Theocharides

Organizations: KIOS Research and Innovation Center of Excellence · Department of Electrical and Computer Engineering, University of Cyprus, 1 Panepistimiou Avenue, 2109 Aglantzia, Nicosia

Abstract

Class imbalance in LiDAR point clouds poses challenges for semantic segmentation in autonomous navigation and urban mapping. While 2D vision has numerous mitigation techniques, their effectiveness in 3D remains unclear. We benchmark six reweighting schemes and five imbalance-aware losses across three datasets (DALES, S3DIS, STPLS3D) using two architectures (KPConv, RandLA-Net). Inverse-frequency weighting degrades performance by up to 12% compared to uniform weighting, with catastrophic failures in minority classes. Uniform weighting performs within 2% of complex losses for structured sampling (KPConv) but benefits less for random sampling (RandLA-Net, up to 4.6% gap). Loss landscape analysis reveals a complex interplay: for structured sampling, imbalance ratio determines landscape geometry on real LiDAR data but decouples from it on synthetic data; for random sampling, landscapes show high sensitivity to dataset geometry regardless of imbalance ratio. For the two evaluated point-based architectures, these results suggest that the interaction between sampling strategy (structured vs. random), imbalance severity, and data acquisition characteristics shapes which mitigation approaches are effective.

Figures & tables

Method	DALES		S3DIS		STPLS3D
Method	KPConv	RandLA	KPConv	RandLA	KPConv	RandLA
uni	80.014	76.762	63.120	61.375	57.093	53.464
invf	67.770	67.372	63.685	60.373	46.874	45.628
cb	78.566	77.068	63.530	62.406	58.194	51.650
invl	80.682	76.843	63.704	61.661	59.284	48.526
invp	80.813	78.481	62.953	63.686	56.430	56.607

Table 1: mIoU (%) per dataset and architecture. Best in bold , failures (

\geq

5% below uniform) in red .

Panel of Figure 1 — Figure 1: Class-wise IoU differences (‘method‘ - ‘uniform‘) of different methods for DALES (upper panel), and zoom-in on smaller classes (lower panel), using KPConv.

Figure 2: Class-wise IoU differences (‘method‘ - ‘uniform‘) of different methods for DALES using RandLA-Net.

Panel of Figure 3 — Figure 3: Class-wise IoU differences of different methods for S3DIS using KPConv (upper panel) and RanLA-Net (lower panel).

Panel of Figure 4 — Figure 4: Class wise IoU differences of different methods for STPLS3D using KPConv (upper panel) and RanLA-Net (lower panel).

Figure 5: Loss landscape flatness analysis for KPConv on DALES, S3DIS, and STPLS3D. The plots show the loss deviation

\delta

as weights are perturbed with magnitude

\alpha

Panel of Figure 6 — Figure 6: Loss landscape flatness analysis for RandLA on DALES, S3DIS, and STPLS3D. The plots show the loss deviation $\delta$ as weights are perturbed with magnitude $\alpha$ . Figure insets (red boxes) provide a zoomed view of low perturbation magnitudes.

Method	ceiling	floor	wall	beam	column	window	door	chair	table	bookcase	sofa	board	clutter
uni	93.636	98.515	80.843	0.000	22.259	44.236	61.458	87.107	79.278	70.984	64.165	60.649	57.433
invf	92.212	98.305	80.333	0.000	25.148	46.485	61.312	87.132	79.381	71.166	70.813	60.570	55.048
cb	92.728	98.405	80.571	0.000	24.341	44.645	62.272	87.440	78.390	71.073	67.183	62.709	56.127
invl	93.443	98.486	80.786	0.000	24.058	45.156	59.885	87.778	79.509	71.855	67.775	60.852	58.569
invp	93.010	98.415	80.546	0.000	23.339	46.768	60.452	87.024	78.933	70.266	60.969	61.074	57.592
comf	92.944	98.410	81.089	0.000	21.425	46.484	66.631	87.458	79.543	71.006	63.212	61.433	56.755

Table S5: Per-class IoU (%) for the S3DIS dataset using KPConv.

Method	ceiling	floor	wall	beam	column	window	door	chair	table	bookcase	sofa	board	clutter
uni	93.116	97.028	80.367	0.000	17.390	57.660	36.838	78.216	84.665	55.844	70.799	71.421	54.524
invf	91.363	97.390	78.549	0.000	15.893	60.357	30.560	76.231	85.703	61.124	71.363	64.896	51.420
cb	91.821	97.074	79.904	0.000	26.317	61.150	33.409	78.371	81.896	75.646	70.521	64.509	50.663
invl	92.351	97.355	80.518	0.000	16.287	59.860	39.551	77.654	86.433	60.295	70.575	68.157	52.555
invp	92.722	97.734	80.037	0.000	22.035	59.819	41.352	78.766	86.838	72.210	71.187	72.921	52.304
comf	92.319	97.902	80.974	0.000	24.176	59.874	33.803	78.060	87.450	61.501	71.115	73.440	53.712

Table S6: Per-class IoU (%) for the S3DIS dataset using RandLA-Net.

Class-Imbalanced Learning

Class-imbalanced learning develops classifiers that recognize rare classes without sacrificing performance on common ones.

See all 257 papers

Loss Landscape Topology Reveals Why Simple Baselines are Competitive at 3D Point Cloud Segmentation Under Class Imbalance
Jul 23, 2026Antonis Savva, Christos Kyrkou, Theocharis TheocharidesClass-Imbalanced Learning Loss Landscape
KIOS Research and Innovation Center of Excellence, University of Cyprus, 1 Panepistimiou Avenue, 2109 Aglantzia, Nicosia · Department of Electrical and Computer Engineering, University of Cyprus, 1 Panepistimiou Avenue, 2109 Aglantzia, Nicosia
Rethinking 3D Segmentation from Individual LiDAR Scans: Incidence-Aware Sampling on the SIP Benchmark
Aug 7, 2026Seongyong Kim, Jingdao Chen, Yong Kwon Cho3D Point Cloud Segmentation 3D Semantic Segmentation
Graduate Student, School of Civil and Environmental Engineering, Georgia Institute of Technology, 790 Atlantic Dr., Atlanta GA · Assistant Professor, Department of Computer Science and Engineering, Mississippi State University, 665 George Perry St, Mississippi State, MS 39762 · Professor, School of Civil and Environmental Engineering, Georgia Institute of Technology, 790 Atlantic Dr., Atlanta, GA
Adaptive Augmentation-Aware Latent Learning for Robust LiDAR Semantic Segmentation
Mar 1, 2026Wangkai Li, Zhaoyang Li, Yuwen Pan +33D Point Cloud Segmentation 3D Semantic Segmentation
University of Science and Technology of China · National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory

Jul 23, 2026J/K move · Enter open · S save

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

Antonis Savva, Christos Kyrkou, Theocharis Theocharides

KIOS Research and Innovation Center of Excellence, University of Cyprus, 1 Panepistimiou Avenue, 2109 Aglantzia, Nicosia · Department of Electrical and Computer Engineering, University of Cyprus, 1 Panepistimiou Avenue, 2109 Aglantzia, Nicosia

Class-Imbalanced Learning Loss Landscape 3D Semantic Segmentation

Open paper AskPDF arXiv

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.

Abstract

Class-Imbalanced Learning

Explore similar work

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

Rethinking 3D Segmentation from Individual LiDAR Scans: Incidence-Aware Sampling on the SIP Benchmark

Adaptive Augmentation-Aware Latent Learning for Robust LiDAR Semantic Segmentation