New paper in IJGIS!

With Shivangi Srivastava, John Vargas and Devis Tuia, we recently published an article on land-use classification using ground based pictures (e.g. Google Street View images):

Fine-grained landuse characterization using ground-based pictures: a deep learning solution based on globally available data

Shivangi Srivastava, John Vargas, Sylvain Lobry, Devis Tuia

in International Journal of Geographical Information Science (IJGIS)

Good news: it is published in open access!

GSV Images

Ground-based pictures (both inside and outside) from Google Street View.

This work assigns a land-use class (e.g. religious place, hotel, post office, …) to buildings, which is an essential task towards better urban land management. In order to do this classification, we use the images from Google Street View pointing to a building as an input to a variable input siamese network (allowing to take into account the information coming from an unknown number of images). This network has been trained using ground truth labels extracted from OpenStreetMap (which are freely available).

PhD award of the Fondation Mines-Télécom

I have been awarded best PhD thesis from the Fondation Mines-Télécom!

Every year, a prize honors the best PhD thesis from the Futur & Ruptures program (co-funded by the Fondation Mines-Télécom and the Carnot institute Télécom & Société numérique).

 

After presenting the work I have done during my PhD at the Journée Futur & Ruptures (a day organized every year by the Fondation Mines-Télécom to present the work achieved during the PhD they funded), four of us were selected as finalists for the PhD award.

I then made a video (of 3 minutes, in french) summing up my work (long way to go before being a youtuber..!). The prizes were announced during the special event organized by the Fondation, which was held on the 27th of March at the BNP Parisbas headquarters in Paris.

Credit: Franck Beloncle

 

The awards were given as follow:

1st prize: Sylvain Lobry – Markovian models for information extraction in SAR imagesTélécom ParisTech

2nd prize (tied):

El Mehdi Amhoud – Coding Techniques for Space Division Multiplexed Optical Fiber Systems – Télécom ParisTech

Yousra Bekhti – Source localization for functional brain imaging with M/EEG – Télécom ParisTech

Ivan Gorynin – Fast Filtering and Unsupervised Estimation in Switching Markov Models Télécom SudParis

It was a great evening, and I particularly enjoyed the interactions with the other recipients! Once again, I would like to thank the Fondation for this prize, as well as Loïc Denis, Florence Tupin, Roger Fjørtoft and Cyrielle Flosi for their advices on the video.

Posted in PhD

PhD defense – November 16th, 2017

On the 16th of November I defended my PhD at Télécom ParisTech. The title is:

Markovian models for SAR images:

Application to water detection in SWOT satellite images and multi-temporal analysis of urban areas

In front of the jury:

You can find more information (including manuscript, slides, animations…) on the dedicated page.

I wanted to thank everyone who made this day feel really special, starting with the member of the jury. Thanks to their challenging questions, the defense was a moment that I enjoyed!

And I would like to thank my colleagues, friends and family who came and organized the traditional cocktail. It was great seeing all of you!

 

Posted in PhD

IGARSS 17

I just came back from IGARSS 17 which took place at Fort Worth, Texas, USA. I did two oral presentations on algorithms dedicated to Water detection in SWOT images:

  • The first one is dedicated to the detection of large water bodies and uses two Markov Random Fields iteratively: the first one do the actual classification, while the second ones estimates the parameters of the class distributions.
  • The other algorithm detects thin water surfaces (such as rivers). It combines a pixel-based detector (which detects parts of the river network) and a higher-level step connecting detected segments.

More information can be found in the proceedings posted in the publication page.