Deep convolutional features for image retrieval

Nowadays, the use of Convolutional Neural Networks (CNNs) has led to tremendous achievements in several computer vision challenges. CNN-based image retrieval methods vary in complexity, growing capacity, and execution time. This work presents a state-of-the-art review in Deep Convolutional Features for image retrieval, pointing out their scope, advantages, and limitations. Moreover, the paper presents a procedure that adopts the latest architectures of pre-trained CNNs that have been initially proposed for image classification to shape image retrieval features. It investigates their suitability on several image retrieval tasks, without any optimization procedure, exhaustive preparatory work, and tuning. Each network's performance is evaluated in two different setups: one employing global and one using local representations. Extensive experiments on several well-known benchmark datasets demonstrate that a simple normalization on the pre-trained networks yields results comparable to state-of-the-art approaches. The global descriptor shapes a plug-and-play approach, which can be adopted for description and retrieval without any prior initialization or training. Moreover, the descriptor's localized version outperforms significantly much more sophisticated and complex methods of the recent literature.