STC: Student-Teacher Collaborative Model for Multi-Target Domain Adaptation

Cross-domain object detection poses significant challenges due to the susceptibility of object detection models to data variance, particularly the domain shifts that can occur between different domains. To address the limitation, we draw inspiration from knowledge distillation, proposing a collaborative learning framework. Our method employs CycleGAN to generate target-style images, and during pretraining, an unsupervised domain adaptation teacher model is trained for each source-target pair. In the distillation process, our proposed algorithm implements an out-of-distribution estimation strategy to select samples that best align with the current model, thereby enhancing the cross-domain distillation process. Furthermore, each expert model is encouraged to collaborate by designating the student model as a bridge between different target domains, facilitated by the Exponential Moving Average (EMA) algorithm. Experiments show that the proposed method leverages structured information, not only does it perform well across various target domains, but it also yields favorable results compared to state-of-the-art unsupervised methods that are specifically trained on single source-target pair.

1. Introduction

Object detection models for driving scenarios rely on large-scale annotated datasets. And the model's performance can significantly deteriorate when evaluated on a dataset with backgrounds or styles that differ from the training data due to model's limited generalization capabilities. Therefore, investigating the model's transfer performance across various scenarios such as the transition between driving conditions collected from different devices and the adaptation to diverse weather conditions is essential for enhancing the model's robustness and generalization.