Research map / Point clouds / Segmentation and detection

An End-to-End Robust Point Cloud Semantic Segmentation Network with Single-Step Conditional Diffusion Models

CVPR 2025 · Point clouds › Segmentation and detection · approach: Point cloud semantic segmentation · 15 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 CDSegNet, an end-to-end point cloud semantic segmentation network using a Conditional-Noise Framework of DDPMs where a noise network produces perturbation features fused into a dominant conditional segmentation network, enabling single-step inference.

Problem

Conditional DDPMs with the Noise-Conditional Framework are slow (many training and inference steps) and hard to fit for 3D scene semantic segmentation, while raw point clouds are noisy and sparse, hurting non-diffusion segmentation methods.

Approach

Swap the roles of the noise and conditional networks: the diffusion-trained Noise Network acts as a lightweight noise-feature generator that perturbs the features of a dominant Conditional Network via a cross-attention Feature Fusion Module, so the segmentation network is a non-diffusion model needing only one inference step.

Key results

• CDSegNet reaches 77.9 mIoU on ScanNet (vs 77.6 PTv3) and 36.3 mIoU on ScanNet200 (vs 35.3 PTv3).
• On nuScenes validation CDSegNet achieves 81.2 mIoU, versus 80.3 for PTv3; on the nuScenes test set it gets 82.0 vs PTv3's 81.2.
• Under Gaussian noise with tau=0.5 on ScanNet, CDSegNet achieves 57.0 mIoU versus 45.8 for PTv3.

Measured effects reported in the paper

Surprising findings

• Removing the diffusion modeling from the noise network (Ours-CN) causes a drop in performance, and the authors attribute this to overfitting, while adding noise-feature perturbations improves generalization.
• CDSegNet performs slightly worse on Poisson noise than baselines at larger noise levels, suggesting robustness depends on closeness of noise distribution to the modeled one.

Limitations

• Performance on Poisson noise is weaker than baselines, indicating robustness is limited to noise distributions close to the modeled one.
• The noise schedule range must be tuned per dataset, trading off performance against noise robustness.
• The noise-fitting contribution to robustness is only conjectured and left for future validation.

Experimental setup

Datasets: ScanNet ScanNet200 nuScenes ModelNet40

Benchmarks: ScanNet ScanNet200 nuScenes ModelNet40

Metrics: mIoU mAcc allAcc accuracy under noise (tau) inference time memory

Baselines: PTv1 MinkUNet ST OctFormer PTv2 PTv3 RangeNet++ PolarNet SalsaNext AMVNet Cylinder3D PVKD

Builds on: DDPM Point Transformer V3 Point Transformer V2 Geometric Loss Strategy Lovasz loss

Closest papers by meaning

• Density-Guided Semi-Supervised 3D Semantic Segmentation with Dual-Space Hardness Sampling (CVPR 2024 · similarity 0.88)
• Rethinking Few-shot 3D Point Cloud Semantic Segmentation (CVPR 2024 · similarity 0.86)
• BFANet: Revisiting 3D Semantic Segmentation with Boundary Feature Analysis (CVPR 2025 · similarity 0.86)
• Diffusion-Driven Two-Stage Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2025 · similarity 0.86)
• PDF: A Probability-Driven Framework for Open World 3D Point Cloud Semantic Segmentation (CVPR 2024 · similarity 0.86)
• LogoSP: Local-global Grouping of Superpoints for Unsupervised Semantic Segmentation of 3D Point Clouds (CVPR 2025 · similarity 0.86)