Babenko et al. demonstrate the usage of deep convolutional neural networks, based on the architecture by Krizhevsky et al. , for image retrieval. They report promising results, especially when re-training networks on appropriate datasets and using different compression techniques. Unfortunately, the implementation as well as the dataset for re-training are not publicly available - merely a list (in Russian) corresponding to the keywords used for the Yandex search engine is provided (see here). However, Babenko et al. claim that a custom version of the original implementation (available here) by Krizhevsky et al. is used. The effectiveness of different layers for image retrieval is compared to several state-of-the-art approaches [2,3,4,5].
What is your opinion on the summarized work? Or do you know related work that is of interest? Let me know your thoughts in the comments below: