Research map / Point clouds / Segmentation and detection

D^3CTTA: Domain-Dependent Decorrelation for Continual Test-Time Adaption of 3D LiDAR Segmentation

CVPR 2025 · Point clouds › Segmentation and detection · approach: LiDAR segmentation for driving · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes D3CTTA, a backpropagation-free online continual test-time adaptation method for LiDAR segmentation that combines distance-aware prototypes with KNN pseudo-label filtering and domain-specific feature decorrelation.

Problem

Continual test-time adaptation of pre-trained LiDAR segmentation models to shifting domains (sensor, weather, corruption) without source data, avoiding costly backpropagation and large batch sizes.

Approach

Builds prototypes per distance-based density region, filters pseudo-labels with KNN spatial consistency, and applies RanPAC-style closed-form feature decorrelation with separate Gram matrices per domain, where domains are detected via BN-statistic similarity and Wasserstein distance.

Key results

• On Synth4D to SemanticKITTI-C, average mIoU of 30.82 vs 24.25 for source and 29.15 for EATA (best prior), a 5.7% gain over previous SOTA, without backpropagation.
• On Synth4D to nuScenes-C, average mIoU of 24.15 vs 20.75 for source and 23.76 for EATA.
• Decorrelation alone raises mean mIoU from 24.25 to 30.17 in the SemanticKITTI-C ablation.

Measured effects reported in the paper

Surprising findings

• Stricter entropy or probability pseudo-label selection lowered mIoU despite higher pseudo-label accuracy, attributed to biased prediction distributions under large domain gaps.
• Pseudo-label accuracy did not directly correlate with final performance; selecting only ground-truth-correct labels did not improve results in one experiment.

Limitations

• Degrades in certain domains (e.g., beam missing, cross sensor on nuScenes-C) and underperforms some baselines there.
• Relies only on a simple distance prior rather than richer geometric features.
• Pseudo-label bias remains an open problem; more refined filtering is needed.

Experimental setup

Datasets: Synth4D SemanticKITTI-C nuScenes-C

Benchmarks: Synth4D to SemanticKITTI-C Synth4D to nuScenes-C repeated domains Synth4D to SemanticKITTI-C

Metrics: mIoU

Baselines: Source TENT GIPSO EATA ViDA T3A PSLabel

Builds on: RanPAC T3A CoTTA GIPSO TENT

Closest papers by meaning

• MOS: Model Synergy for Test-Time Adaptation on LiDAR-Based 3D Object Detection (ICLR 2025 · similarity 0.88)
• An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 2024 · similarity 0.87)
• Construct to Associate: Cooperative Context Learning for Domain Adaptive Point Cloud Segmentation (CVPR 2024 · similarity 0.86)
• Multi-Space Alignments Towards Universal LiDAR Segmentation (CVPR 2024 · similarity 0.86)
• TREND: Unsupervised 3D Representation Learning via Temporal Forecasting for LiDAR Perception (NeurIPS 2025 · similarity 0.86)
• Three Pillars Improving Vision Foundation Model Distillation for Lidar (CVPR 2024 · similarity 0.86)