Training & Label-Free Domain Adaptation in 3D Object Detection for Autonomous Driving

Arjun Singh, Arnab Kumar Mondal, Udit Sharma, Saswati Dana, Prathosh AP, Dinesh Garg, Amith Singhee

Keywords: 3D Object Detection, Domain Adaptation, Domain Generalization, Test Time Adaptation, Autonomous Driving

Abstract: Autonomous driving systems rely on 3D object detection for path planning and control, but it has been demonstrated that LiDAR-based 3D object detectors suffer a significant performance drop when tested on datasets from different geographic locations than their training data due to domain shift. Existing domain adaptation methods typically require access to original training data or substantial target domain data, making them impractical for resource-constrained deployments. In this work, we address this challenge by exploring how LiDAR-based 3D object detection models trained in a source domain can be reused in a data-scarce target domain without extensive fine-tuning. We propose lightweight and practical approaches for single-frame domain adaptation under three different resource-constraint settings: 1) a test-time approach that adjusts predictions using a calibration factor when only *a minimal amount of unlabeled target data* is available, 2) a lightweight adaptation technique using learnable scaling weights when *a small amount of labeled target data* are available, while preserving original model parameters, and 3) a weakly supervised finetuning method with a novel loss function for scenarios with *larger amounts of unlabeled target domain data*. Our results demonstrate that scalable, cross-regional deployment of 3D object detection systems is feasible under significant scarcity of data and training budget.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 9111