Research map / Detection and segmentation / 3D detection and driving

An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains

CVPR 2024 · Detection and segmentation › 3D detection and driving · approach: Depth-based and label-free 3D detection · 7 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.

Empirical study of how architecture, voxel height, augmentations, and anchor size affect the out-of-domain robustness of nine Lidar 3D object detectors across six benchmarks covering sensor, weather, and location gaps.

Problem

Lidar 3D object detectors degrade substantially on unseen domains (different sensors, weather, locations), and existing domain adaptation work treats detectors as black boxes without studying how design choices affect robustness before adaptation.

Approach

Systematically isolates four design choices (architecture, voxel encoding, data augmentation, anchor strategy) in OpenPCDet-based detectors and measures their effect on cross-domain performance, rather than proposing a new adaptation method; key levers include test-time anchor resizing, higher voxel height, line/shape downsampling augmentations, and training on clean-weather data.

Key results

• Test-time anchor size tuning substantially raises target AP without retraining, e.g. SECOND on Waymo-to-KITTI car AP rises from 16.41 to 66.81 (Table 5).
• Line downsampling raises PVRCNN K64→K16 car AP from 54.07 to 71.57 and pedestrian AP from 27.13 to 51.64 (Table 6).
• Models trained on clean Waymo data match or exceed oracle models trained on Kirkland bad-weather data on the Kirkland validation split, especially for pedestrians (Figure 7).

Measured effects reported in the paper

Surprising findings

• Models trained only on clean-weather Waymo data perform comparably to or better than the in-domain oracle trained on bad-weather Kirkland data.
• Test-time anchor size adjustment, which tunes on labeled target data, yields large gains, and the best test anchors are always smaller than the training anchor.

Limitations

• Test-time anchor tuning uses labeled target data, so it is not a true domain generalization method and is limited to semi-supervised or few-label settings.
• Greedy anchor search is heuristic and may not find the optimal anchor.
• Waymo training uses only a 20% subset to limit compute, and experiments focus on car for location gap.

Experimental setup

Datasets: KITTI Waymo nuScenes Kirkland

Benchmarks: Waymo-to-KITTI Waymo-to-nuScenes nuScenes-to-KITTI KITTI64-to-KITTI32 KITTI64-to-KITTI16 Waymo-to-Kirkland

Metrics: 3D AP (40 recall points) mAP

Baselines: PointRCNN PointPillars SECOND VoxelRCNN VOTR-VoxelRCNN CenterPoint PVRCNN-Centerhead PVRCNN VOTR-TSD SPG

Builds on: OpenPCDet ST3D SN SPG Lidar beam downsampling (Lidar beam distillation)

Closest papers by meaning

• MOS: Model Synergy for Test-Time Adaptation on LiDAR-Based 3D Object Detection (ICLR 2025 · similarity 0.90)
• GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection (CVPR 2025 · similarity 0.88)
• CMDA: Cross-Modal and Domain Adversarial Adaptation for LiDAR-Based 3D Object Detection (AAAI 2024 · similarity 0.88)
• L4DR: LiDAR-4DRadar Fusion for Weather-Robust 3D Object Detection (AAAI 2025 · similarity 0.88)
• Towards Robust 3D Object Detection with LiDAR and 4D Radar Fusion in Various Weather Conditions (CVPR 2024 · similarity 0.88)
• Training an Open-Vocabulary Monocular 3D Detection Model without 3D Data (NeurIPS 2024 · similarity 0.88)