AFAN: Augmented Feature Alignment Network for Cross-Domain Object Detection

Unsupervised domain adaptation for object detection is a challenging problem with many real-world applications. Unfortunately, it has received much less attention than supervised detection. Models that try to address this task tend suffer from shortage of annotated training samples. Moreover, existing methods feature alignments are not sufficient learn domain-invariant representations. To these limitations, we propose novel augmented alignment network (AFAN) which integrates intermediate image generation and domain-adversarial into unified framework. An generator proposed enhance by automatically generated soft labels. The synthetic images progressively bridge the divergence augment source data. A pyramid designed corresponding discriminator used align multi-scale convolutional features different semantic levels. Last but least, introduce region an instance proposals. Our approach significantly outperforms state-of-the-art on standard benchmarks both similar dissimilar adaptations. Further extensive experiments verify effectiveness each component demonstrate can