Semantics-Guided Contrastive Network for Zero-Shot Object detection.

Zero-shot object detection (ZSD), the task that extends conventional detection models to detecting objects from unseen categories, has emerged as a new challenge in computer vision. Most existing approaches on ZSD are based on a strict mapping-transfer strategy that learns a mapping function from visual to semantic space over seen categories, then directly generalizes the learned mapping function to unseen object detection. However, the ZSD task still remains challenging, since those works fail to consider the two key factors that hamper the ZSD performance: (a) the domain shift problem between seen and unseen classes leads to poor transferable ability of the model; (b) the original visual feature space is suboptimal for ZSD since it lacks discriminative information.To alleviate these issues, we develop a novel Semantics-Guided Contrastive Network for ZSD (ContrastZSD), a detection framework that first brings the contrastive learning paradigm into the realm of ZSD. The pairwise contrastive tasks take advantage of class label and semantic relation as additional supervision signals. Under the guidance of those explicit semantic supervision, the model can learn more knowledge about unseen categories to avoid over-fitting to the seen concepts.