Domain adaptive object detection with generative adversarial network

Domain adaptive object detection has emerged to alleviate the lack of massive amounts of labeled data in the target field. Currently, Generative Adversarial Network (GAN) has been exhibited its benefits to the Faster R-CNN for some specific application. However, it brings up an open question that GAN still has flexibility and robustness in more generic scenarios. Thus, this paper aims to extensively evaluate its efficiency of the GAN on more generic domain-adaptive object detection. More specifically, it presents four generic domain-adaptive object detection with GAN architectures, containing Faster R-CNN with CycleGAN, Faster R-CNN with DualGAN, RetinaNet with CycleGAN, and RetinaNet with DualGAN. Extensive experiment results demonstrate that GAN-derived model makes significant contribution and consistent improvements in two domain shift scenarios for domain-adaptive object detection. Additionally, comparing with the state-of-the-art methods, domain-adaptive object detection with GAN has been validated its ability of generalization and robustness in more challenging scenarios.