Instance Retrieval Using Region of Interest Based CNN Features

Abstract

: Recently, image representations derived by convolutional neural networks (CNN) have achieved promising performance for instance retrieval, and they outperform the traditional hand-crafted image features. However, most of existing CNN-based features are proposed to describe the entire images, and thus they are less robust to background clutter. This paper proposes a region of interest (RoI)-based deep convolutional representation for instance retrieval. It first detects the region of interests (RoIs) from an image, and then extracts a set of RoI-based CNN features from the fully-connected layer of CNN. The proposed RoI-based CNN feature describes the patterns of the detected RoIs, so that the visual matching can be implemented at image region-level to effectively identify target objects from cluttered backgrounds. Moreover, we test the performance of the proposed RoI-based CNN feature, when it is extracted from different convolutional layers or fully-connected layers. Also, we compare the performance of RoI-based CNN feature with those of the state-of-the-art CNN features on two instance retrieval benchmarks. Experimental results show that the proposed RoI-based CNN feature provides superior performance than the state-of-the-art CNN features for in-stance retrieval.