Semi-Supervised Diversity-Aware Domain Adaptation for 3D Object detection

TL;DR

The paper tackles cross-domain generalization for LiDAR-based 3D object detection in autonomous driving by introducing a neuron-activation-pattern–driven Diverse Frame Selection method that enables annotation-efficient domain adaptation. It couples Source–Target Distribution Alignment (point density and bounding-box size) with a frame-selection algorithm that extracts activation patterns from an ROI Head layer, optimizing frame diversity through entropy and inter-frame distance measures, and complements this with post-training strategies to limit weight drift. Empirical results on KITTI, NuScenes, and Waymo show that annotating as few as 10 target frames, chosen for diversity, can outperform linear probing and approach or surpass state-of-the-art methods, albeit with some degradation on the source domain after adaptation. The approach offers a practical pathway for OEMs to extend the operational design domain to new regions or sensor setups with limited labeling, while highlighting the domain-specific nature of adaptation and the value of continual-learning–aware fine-tuning.

Abstract

3D object detectors are fundamental components of perception systems in autonomous vehicles. While these detectors achieve remarkable performance on standard autonomous driving benchmarks, they often struggle to generalize across different domains - for instance, a model trained in the U.S. may perform poorly in regions like Asia or Europe. This paper presents a novel lidar domain adaptation method based on neuron activation patterns, demonstrating that state-of-the-art performance can be achieved by annotating only a small, representative, and diverse subset of samples from the target domain if they are correctly selected. The proposed approach requires very small annotation budget and, when combined with post-training techniques inspired by continual learning prevent weight drift from the original model. Empirical evaluation shows that the proposed domain adaptation approach outperforms both linear probing and state-of-the-art domain adaptation techniques.

Figures (8)

• Figure 1: Datasets. Exemplary databases captured in different ODDs. They utilize diverse LiDAR sensors, employ varied labeling specifications, and have different regions annotated. This diversity poses challenges in generalizing for 3D object detection models.
• Figure 2: Vehicles of different sizes have different representations in the LiDAR point cloud, resulting in substantially different features being extracted.
• Figure 3: Exemplary object detection in new domain (top) without and with (bottom) domain adaptation.
• Figure 4: Distribution of volumes of vehicle annotations. These distributions result from the cultural and geographical features as well as from the labeling guideline.
• Figure 5: Distribution of LiDAR points in the annotation boxes. These distributions result from the point cloud density and the labeling policy