Unsupervised 3D Object Detection Domain Adaptation Based on Pseudo-Label Variance Regularization

Cross-domain detection frequently encounters a decline in detection accuracy, necessitating the application of domain adaptation techniques. One crucial approach to unsupervised domain adaptation is the pseudo label-based self-training method, which iteratively trains the model by treating the pseudo labels as ground truth. However, differences in distribution that can exist between the source and target domains can lead to potentially incorrect generated pseudo labels. This can result in the threshold-setting method failing to accurately select the pseudo labels. Therefore, to tackle the challenge of determining pseudo label thresholds in self-training, we propose an unsupervised 3D object detection domain adaptation method based on pseudo label regularization. Specifically, a self-training framework based on the fusion of two detection heads is used to obtain more accurate pseudo labels. The variance of the two detection heads is utilized as the noise information for the corresponding pseudo labels. Then, the noise information is incorporated as a regularization term to enhance the bounding box regression loss, thereby addressing the challenge of determining pseudo label thresholds in self-training. The experimental results demonstrate that the method proposed in this paper achieves higher cross-domain detection accuracy compared to existing domain adaptation methods for 3D object detection.