Feature Enhancement for Multi-scale Object Detection

Recently, deep learning has brought great progress in object detection. However, we believe that traditional hand-crafted features may still contain valuable human knowledge complementary to features learned from raw data. Besides, almost all top-performing object detection methods extract features by using backbones originally designed for image classification. The generated features are often highly semantic, which is beneficial to global image classification, but may lose details useful for object localization and recognition under various scales. To alleviate the problems mentioned above, a feature enhancement method is proposed in this paper. Inspired by the success of histograms of oriented gradients in traditional object detection research, we construct feature channels based on oriented gradients as input to convolutional neural networks to capture discriminative local orientations. The oriented gradients and RGB features are stacked as input of network to enhance the input feature representation. For accurate object localization and recognition, we employ dilated convolutions to increase spatial resolutions of output feature maps while maintaining their respective receptive fields. Hierarchical feature maps with different receptive fields are aggregated into the final feature representation for multi-scale object detection without extra upsampling. Experimental results on PASCAL VOC 2007 and 2012 demonstrate superiority of the proposed method compared with state-of-the-art methods for multi-scale object detection.