Image retrieval via CNNs in TensorFlow 2

This thesis addresses the problem of instance-level image retrieval in largescale picture collections, intending to find the greatest number of images corresponding to a query. Convolutional neural networks (CNNs) have demonstrated their ability to provide effective descriptors for content-based image retrieval (CBIR). Given the current knowledge, we focused our efforts on utilizing fine-tuned CNNs for global feature extraction with the goal of using those for image retrieval problems. Firstly, we examined several methods proposed to improve image retrieval, such as GeM [RTC18] and DELF [NAS+17]. As the main result of this thesis, an extendable and highly-customizable image retrieval framework based on the work of Radenović et al. [RTC18] was re-implemented in TensorFlow 2. This approach produces state-of-the-art retrieval results, while using relatively short descriptors. As a validation, we trained the networks on the SfM120k landmark images dataset and performed experiments on two image retrieval benchmarks (revisited Oxford5k and Paris6k). Different training strategies, network architectures and loss functions were used in the experiments. The final project code was successfully merged into the official Tensorflow repository managed by Google, as a part of the DELF [Tena] research library.