Object Detection for Chinese Traditional Costume Images Based GRP-DSOD++ Network

Haiying Zhao, Ting Yang, Xiaogang Hou, Hui Zhu, Zhuoyu Yang

Abstract: The image object detection methods based on deep learning have achieved remarkable results in recent years. However, as object sizes of Chinese Traditional Costume Images (CTCI-4) data set are smaller than that of natural images, and there are not enough training samples, the previous excellent object detection methods cannot achieve good detection result. To tackle this issue, mainly inspired by GRP-DSOD, we propose an effective network, namely GRP-DSOD++ network, to detect objects in the CTCI-4 data set. In order to collect multi-scale context information and capture a wider range of features, we introduce Dilated-Inception module (DI module) and applied it to object detection framework that is learned from scratch. We also applied other advanced components of several excellent object detectors to the proposed network architecture. The proposed detector in the CTCI-4 data set achieves 77.08% mAP, higher than the GRP-DSOD detector (75.33% mAP). And the detector (learning on VOC “07+12” trainval) also can achieve good performance on PASCAL VOC2007.