Landmark Recognition is a large-scale classification task which is raising the interest of machine learners from all over the world for its peculiar challenging aspects. A huge amount of landmarks have been captured and gathered in the Google Landmark Dataset. This extreme classification scenario is combined with the necessity to confirm the correctness of the obtained prediction, since the test set contains ’tricky’ samples. To do so, a retrieval mechanism based on Deep Local Features is implemented. This solution provides a full pipeline composed by three main stages:
- A pre-processing and filtering step to deal with constraints imposed by both time and computational resources available;
- The classification phase has been performed with a ResNet50 model exploiting transfer learning from ImageNet dataset;
- DELF module has been adapted to our specific application with a thresholdbased decision system for an efficient verification of the predictions.
In the following paper the results will be analyzed.