mar. 11 septembre 2018

Hello everyone,

Last July I was in Valencia (Spain) for IGARSS 2018, the main international remote sensing conference with about 3000 registered attendees. I was there to present our paper on generating hyperspectral samples with GANs and to see what the trending topics are for 2019 !

IGARSS 2018 venue

In 2016, IGARSS was held in Beijing where the heavy air kept us very warm. This year in Valencia, the mediteranean breeze cooled us down... a bit. It was hot, with temperatures ranging from 43°C to a mere 29°C at night. The venue had air conditionners although the main floor was located under a huge glass dome which seems like a very poor idea during the summer. Anyway, I don't really have anything to add, it was the straightforward conference center that I expected. My main remark is relevant to the organization committee, that somehow planned all the deep learning and machine learning sessions in the smallest rooms possible, with attendees not able to sit down or even enter the room ! Meanwhile, the main stage with hundreds of seats was desperately empty.

Technical content

As usual, I can only talk about what I've seen, since IGARSS is a huge conference that adresses lots of specific fields and research areas. I mostly focused on remote sensing data interpretation and well... I am critical about was presented this year. Starting with the good things, I attended some very good sessions (mostly invited) about deep learning for various aerial and satellite sensors. A few excellents talks related to principled approaches of deep nets for scene understanding: semantic segmentation from scribbled annotations, cloud removal using GANs and cloud tracking with multi-temporal deep networks. Overall, I was quite happy with several talks that introduced clever ideas to combine state-of-the-art deep learning and remote sensing data.

However, I was disappointed with the quantity of papers where the contribution can be summed up as "we applied model X on dataset Y". This was already the case during IGARSS 2016 but the deep learning hype was huge and most of these experiments were still new for the remote sensing community. Although, this is not the case in 2018 anymore. There are at least three common pitfalls that I've seen shared by lots of papers, both in poster and oral presentations: * no methodological analysis, e.g. "we applied Faster-RCNN on our boat detection dataset". Why ? Would a simpler model work ? Did you perform a robustness assessment ? Doesn't the model overfit to your small dataset ? * wrong train/test split. It is not okay to build a train/test split by doing a random sampling of the pixels, especially if your model is a CNN with a large receptive field. If you want to evaluate the generalization ability of your model, then you must tune your hyperparameters on a validation set and evaluate on a separate, spatially disjoint, test set. See Jordi's Inglada write-up on this topic. * not taking into account what makes remote sensing data special, e.g. using only RGB bands in multispectral data or performing random mirroring and flipping on SAR data. I get it that most of the deep learning for image interpretation papers come from computer vision, but multimedia images and remote sensing data are not the same thing. Multispectral and hyperspectral cameras can see things invisible in RGB. Radar involves complex physical phenomenon and you cannot just throw 2D CNNs at SAR data, because they are not images. I think that these problems will smooth out as more and more people coming from remote sensing steps into machine learning/computer vision (and vice-versa), but it remains frustrating to see so many papers with significant flaws. There are excellent ideas out there but poorly evaluated because of a lack of standardized benchmarks and common methodological processes.

Valencia

Valencia is a nice town to walk around. I liked that the city is very colourful and that you can see all of the historical center on foot. I especially enjoyed visiting the ancient city gates and the Oceanografic. We also visited the science museum, which is more kid-oriented and all in all not as fascinating as the ones in Paris (sorry!). There is however an excellent archeological museum with underground ruins of the roman city that I recommend. The gardens surrounding the city are also a must if you're there.

Until next time !

Category: conference