A key challenge in generic object detection is being to handle large variations in object scale, poses, viewpoints, especially part deformations when determining the location for specified object categories. Recent advances in deep neural networks have achieved promising results for object detection by extending the traditional detection methodologies using the convolutional neural network architectures. In this paper, we make an attempt to incorporate another traditional detection schema, Regionlet into an end-to-end trained deep learning framework, and perform ablation studies on its behavior on multiple object detection datasets. More specifically, we propose a "region selection network" and a "gating network". The region selection network serves as a guidance on where to select regions to learn the features from. Additionally, the gating network serves as a local feature selection module to select and transform feature maps to be suitable for detection task. It acts as soft Regionlet selection and pooling. The proposed network is trained end-to-end without additional efforts. Extensive experiments and analysis on the PASCAL VOC dataset and Microsoft COCO dataset show that the proposed framework achieves comparable state-of-the-art results.

本论文提出了一种名为“Deep Regionlets”的新型目标检测框架，通过在深度神经网络和传统检测模式之间建立桥梁实现准确的普适目标检测。该框架由区域选择网络和深度区域学习模块组成，能够实现端到端的训练，并通过单元素件选择和转换减轻局部变化。此框架在PASCAL VOC和Microsoft COCO数据集上的实验结果表明，其效果优于现有的RetinaNet和Mask R-CNN算法。

深度区域特征检测