Project E/R on Visual Landmark Recognition

The project is totally funded by the italian region Emila Romagna.

The project has the aim of studying and implementing algorithms of image processing and pattern recognition for the analysis of large datasets of images and video to identify automatically points, objects and buildings of interest (landmarks) in urban or industrial environment.
By comparing the images and videos captured by mobile devices (cameras mounted on vehicles) or from devices such as smartphones and tablets with an image database of places and known or recognizable objects, the system must be able to geo-locate the user (person or vehicle).

An important feature (and distinctive compared to existing experiences) is the process of “population” of visual databases that will be collaborative and incremental.


The landmarks (and then the images that represent them to be included in the database) can be added by users in a kind of collaborative social networks.

This, on the one hand, ensure a greater amount of data available for the recognition (by increasing the size of the visual Big Data) and, on the other hand, will generate a series of technological and scientific challenges of great interest for the academic community and industrial.

The fields of application of these techniques range from geoocalizzazione in urban environments for navigation systems and recommendation, the location of vehicles and people in industrial environments (intelligent logistics systems), a generic image retrieval applications for the recognition of objects and places.

Federico Magliani is the supervisor of this project.

The steps executed:

  • an improvement of an image dataset: ZuBuD+
  • a creation of a new algorithm for image retrieval problem: locVLAD, that outperforms the state-of-the-art on intra-dataset problem in most famous public datasets such as ZuBuD and Holidays
  • the use of “transfer learning” for improving the retrieval performance, using locVLAD descriptors. It consists in tuning the parameters trained in one feature space in order to work in another feature space
  • the implementation of a family of algorithms of indexing methods called Bag of Indexes for large-scale recognition based on Locality-Sensitive Hashing (LSH) and its variants which allows to minimize the accuracy reduction with the growth of the data.


If you used this dataset please cite the following paper: A location-aware embedding technique for accurate landmark recognition.


  title={A location-aware embedding technique for accurate landmark recognition},
  author={Magliani, Federico and Bidgoli, Navid Mahmoudian and Prati, Andrea},
  journal={arXiv preprint arXiv:1704.05754},

Now we are working on inter-dataset problem, looking for a solution, that obtains good results also in mobile devices (reduced size of descriptors and computation).

PDF Poster for ICVSS 2017