Investigate an Intelligent Vehicle Domain Adaptation Object Detection Algorithm Based on Self-Training and Adversarial Training

There are evident data distribution differences between the source domain and the target domain, posing challenges such as domain shift and label space shift in domain adaptive object detection. Currently, mainstream methods employ differential and adversarial strategies to achieve domain alignment. Differential strategies suffer from issues like high computational complexity and sensitivity to hyperparameters, while adversarial strategies are prone to interference from background noise and object position variations.Furthermore, some approaches attempt domain adaptation based on these strategies, such as using pseudo-labels for self-training. However, self-training methods often rely on high-confidence predictions as pseudo-labels in unlabeled domains. Due to potential calibration issues during domain transfer, the accuracy of these high-confidence predictions cannot be guaranteed. This paper proposes a collaborative framework based on self-training and adversarial training, guided by uncertainty, achieving a balance between adversarial feature alignment and self-training. Extensive experimental validation across various domain adaptation scenarios demonstrates the effectiveness of our framework on anchor-free FCOS detectors with different backbone networks compared to existing algorithms.