Curriculum self-paced learning for cross-domain object detection

Training (source) domain bias affects state-of-the-art object detectors, such as Faster R-CNN, when applied to new (target) domains. To alleviate this problem, researchers proposed various adaptation methods improve detection results in the cross-domain setting, e.g. by translating images with ground-truth labels from source target using Cycle-GAN. On top of combining Cycle-GAN transformations and self-paced learning a smart efficient way, paper, we propose novel algorithm that learns easy hard. Our method is simple effective, without any overhead during inference. It uses only pseudo-labels for samples taken domain, i.e. unsupervised. We conduct experiments on four benchmarks, showing better than state art. also perform an ablation study demonstrating utility each component our framework. Additionally, applicability framework other detectors. Furthermore, compare difficulty measure measures related literature, proving it yields superior correlates well performance metric.