sam. 05 août 2017

Hello everyone,

Last July I was far, far away from home since I was in Honolulu, Hawaii for the CVPR 2017 conference ! I count myself lucky to have this kind of opportunities during my PhD thesis, rest assured. Anyway, my presence there was motivated by the fact that I was presenting our work on fusing OpenStreetMap data and aerial images to perform state-of-the-art semantic labeling of urban areas. As usual, I detailed some of my thoughts below.

CVPR 2017 venue

So, compared to my previous conferences, CVPR was somewhat intimidating. First, it is a huge conference with more than 5000 registrations (and probably even more if you cont late on-site registrations). Second, as it is one of the most prestigious conference in Computer Vision, there were lots of great researchers walking around. Therefore, I felt sometimes a bit out-of-place in this huge high-level crowd. However, the venue was nice, with excellent food, nice chairs with big screens for slides and very helpful volunteers. Plus, Hawaii is obviously an amazing place to chill in !

Technical content

Lots of things to see at CVPR ! I mostly focused my attention on deep learning related to semantic segmentation and/or semi-supervised learning, but I wandered around during the poster sessions. I presented during the CVPR EarthVision workshop which was very interesting and I met lots of people that asked insightful questions about why and how we use deep learning for cartography. During the oral talks, I liked the spotlight format that quickly allowed to check whether I was interested in a specific paper, although it is challenging for the speaker to keep it short (<5 minutes).

The plenary sessions were adequate. I thoroughly enjoyed James DiCarlo's take on why and how to bridge the gap between machine learning and neurosciences. On the other hand, Dan Jurasky talked about languages and language processing, which is in itself interesting but not really what I expected from a computer vision conference. And Harry Shum's presentation was basically a one-hour long advertisement for the HoloLens. Oh, well... The regular oral sessions were okay.

On the science itself, I feel that many papers were in the "soft-belly" of the curve : interesting but not really ground-breaking. I saw lots of ideas applied to deep learning that were mostly some weird tricks to slightly boost the accuracy on dataset X or Y, but few that I would actually use. Moreover, many good papers were already available on arXiv. However, there was some great papers that gave me new ideas and that I had not seen before, especially stuff relative to weakly supervised semantic segmentation, aerial images processing and fun architectures such as DenseNets.

Honolulu

Honolulu is the biggest city in the Hawaii archipel. Waikiki is really a sight to behold with its tiki-style hotels, the amazingly blue sea and the sunny weather. I got around the city a bit, to see the zoo, the aquarium but more importantly, the Diamond Head crater and Pearl Harbour. From the Diamond Head, there is an incredible view of the coast and Honolulu itself and the hike is definitely worth it. Pearl Harbour was interesting, although I think the memorial part did not resonate with me as much as it probably resonates in the heart of the American tourists that were there. The Bishop Museum is also worth the detour if you want to know more about the Hawaiian and Polynesian culture and history. I also did the planetarium show, because who doesn't like to look at stars ?

Until next time !

Category: conference