Why are structured light sensors not used instead of color camera images? Because in sunlight it becomes impossible to see your own projected IR light. So they stop working.
And projecting a 3D 360 degrees point cloud from a LIDAR to a 2D image is weird. Why take that loss when the algorithms can just work on the point cloud.
>> This class of low-cost depth sensors emerged just as Deep Learning boosted Artificial Intelligence by accelerating performance of hyper-parametric function approximators leading to surprising advances in image classification, speech recognition, and language translation.
So I haven't really looked at machine translation benchmarks for a while, but the paper linked by the article itself (as an example of the "surprising advances" in language translation) reports results below the state of the art (from 2014) (see [1] below).
The system noted as the state of the art, [2], is a phrase-based statistical machine translation system, which basically boils down to n-grams, with a few tricks up their sleeves such as a 5-gram operation sequence model, hierarchical lexicalised reordering, an unsupervised transliteration model (for translations between Hindi and Urdu) etc etc.
This is interesting to me because I hear this claim, that deep learning has advanced machine translation significantly, very often- but when I look for results, I can't really find anything that supports that. It's obvious that machine translation research currently involves a lot of deep learning, but whether that is actually working, or it's just what people are trying at the moment to see if it works, I can't tell.
It certainly doesn't look -to me at least- like we've seen anything like the big leap on ImageNET and CIFAR results, when it comes to machine translation. The paper linked by the article is a good example: their baseline is a 33.30 BLEU score and their system raises that to 34.81, an improvement of 1.51, a less than 5% increase (on the previous result). And that's only on English-to-French (and still behind the top results).
I really don't see the "surprising advancements" here.
And projecting a 3D 360 degrees point cloud from a LIDAR to a 2D image is weird. Why take that loss when the algorithms can just work on the point cloud.