Feature pyramids have delivered significant improvement in object detection. However, building effective feature pyramids heavily relies on expert knowledge, and also requires strenuous efforts to balance effectiveness and efficiency. Automatic search methods, such as NAS-FPN, automates the design of feature pyramids, but the low search efficiency makes it difficult to apply in a large search space. In this paper, we propose a novel search framework for a feature pyramid network, called AutoDet, which enables to automatic discovery of informative connections between multi-scale features and configure detection architectures with both high efficiency and state-of-the-art performance. In AutoDet, a new search space is specifically designed for feature pyramids in object detectors, which is more general than NAS-FPN. Furthermore, the architecture search process is formulated as a combinatorial optimization problem and solved by a Simulated Annealing-based Network Architecture Search method (SA-NAS). Compared with existing NAS methods, AutoDet ensures a dramatic reduction in search times. For example, our SA-NAS can be up to 30x faster than reinforcement learning-based approaches. Furthermore, AutoDet is compatible with both one-stage and two-stage structures with all kinds of backbone networks. We demonstrate the effectiveness of AutoDet with outperforming single-model results on the COCO dataset. Without pre-training on OpenImages, AutoDet with the ResNet-101 backbone achieves an AP of 39.7 and 47.3 for one-stage and two-stage architectures, respectively, which surpass current state-of-the-art methods.

特征金字塔已大大改善了对象检测。但是，构建有效的特征金字塔在很大程度上依赖于专家知识，并且还需要付出艰辛的努力来平衡有效性和效率。自动搜索方法（例如NAS-FPN）使要素金字塔的设计自动化，但是搜索效率低，使其难以在较大的搜索空间中应用。在本文中，我们提出了一种用于特征金字塔网络的新颖搜索框架，称为AutoDet，该框架可自动发现多尺度特征之间的信息连接，并以高效和最新性能配置检测架构。在AutoDet中，专门为对象检测器中的特征金字塔设计了一个新的搜索空间，它比NAS-FPN更通用。此外，架构搜索过程被表述为组合优化问题，并通过基于模拟退火的网络架构搜索方法（SA-NAS）解决。与现有的NAS方法相比，AutoDet可确保大大减少搜索时间。例如，我们的SA-NAS可以比基于强化学习的方法快30倍。此外，AutoDet与具有各种骨干网的一级和二级结构兼容。我们在COCO数据集上以优于单一模型的结果证明了AutoDet的有效性。在没有对OpenImages进行预训练的情况下，具有ResNet-101主干的AutoDet对于一级和两级体系结构的AP分别达到39.7和47.3，这超过了当前的最新方法。