Instance-level object retrieval via deep region CNN

Abstract

Instance retrieval is a fundamental problem in the multimedia field for its various applications. Since the relevancy is defined at the instance level, it is more challenging comparing to traditional image retrieval methods. Recent advances show that Convolutional Neural Networks (CNNs) offer an attractive method for image feature representations. However, the CNN method extracts features from the whole image, thus the extracted features contain a large amount of background noisy information, leading to poor retrieval performance. To solve the problem, this paper proposed a deep region CNN method with object detection for instance-level object retrieval, which has two phases, i.e., offline Faster R-CNN training and online instance retrieval. First, we train a Faster R-CNN model to better locate the region of the objects. Second, we extract the CNN features from the detected object image region and then retrieve relevant images based on the visual similarity of these features. Furthermore, we utilized three different strategies for feature fusing based on the detected object region candidates from Faster R-CNN. We conduct the experiment on a large dataset: INSTRE with 23,070 object images and additional one million distractor images. Qualitative and quantitative evaluation results have demonstrated the advantage of our proposed method. In addition, we conducted extensive experiments on the Oxford dataset and the experimental results further validated the effectiveness of our proposed method.