SCAN: Semantic context aware network for accurate small object detection

Abstract

Recent deep convolutional neural network-based object detectors have shown promising performance when detecting large objects, but they are still limited in detecting small or partially occluded ones—in part because such objects convey limited information due to the small areas they occupy in images. Consequently, it is difficult for deep neural networks to extract sufficient distinguishing fine-grained features for high-level feature maps, which are crucial for the network to precisely locate small or partially occluded objects. There are two ways to alleviate this problem: the first is to use lower-level but larger feature maps to improve location accuracy and the second is to use context information to increase classification accuracy. In this paper, we combine both methods by first constructing larger and more meaningful feature maps in top-down order and concatenating them and subsequently fusing multilevel contextual information through pyramid pooling to construct context aware features. We propose a unified framework called the Semantic Context Aware Network (SCAN) to enhance object detection accuracy. SCAN is simple to implement and can be trained from end to end. We evaluate the proposed network on the KITTI challenge benchmark and present an improvement of the precision.