Maxime Lenormand

@MaxLenormand

MSc. in Aerospace Engineering. Finishing . expert. Fighting malaria w. ML & sat imagery w. . Looking for a job!

Somewhere on Earth, for now
Vrijeme pridruživanja: svibanj 2017.

Tweetovi

Blokirali ste korisnika/cu @MaxLenormand

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @MaxLenormand

  1. Prikvačeni tweet
    14. stu 2019.

    Hi Twitter! Nearly graduated, I’m looking for my 1st job! I’m a 23 y.o. , but also a lot into & . I’d love to work in computer vision applied to . Free starting Feb/March 2020, anywhere in the 🌍! RT would be awesome! 🛰

    , , i još njih 5
    Poništi
  2. 31. sij

    I can understand that people might not understand why some of us care about our privacy & go out of our way to prevent our data from being collected. isn't a tech guy, and it's great that more people outside of the field talk about it too:

    Poništi
  3. 23. sij

    After joining , the biggest competitive data science platform out there about a year ago, I finally reached the expert ranking! Learned a lot & keep learning every day on there, awesome tool to keep up to date with the latest research & how to apply it

    Poništi
  4. 22. sij

    Me: Ah damn it I'm so tired I don't feel like writing my internship report, takes too much brain power... Also me: Ooooouuh a really interesting state of the art 15 page paper on deep learning, let's read it multiple times!

    Poništi
  5. 21. sij

    Oh, and I should mention, I'm also looking for a job! If you're looking for someone with machine learning / deep learning / computer vision experience applied to satellite imagery or anything else, you just read what I did for the past ~6 months!

    Prikaži ovu nit
    Poništi
  6. 21. sij

    I hope this was interesting to follow, and actually makes sense. Cutting down into tweets is far from easy, but a nice exercise. If you have any questions, feel free to reach out!

    Prikaži ovu nit
    Poništi
  7. 21. sij

    Again, I won't go into detail about how this is done. For those interested, we use a UNet (done on Keras by the way, poke ). This is the result (trained on Ghana, deployed in Nigeria).

    Prikaži ovu nit
    Poništi
  8. 21. sij

    I also ended up being the person with the most experience in computer vision using deep learning (those longs nights on have been useful) in the company. So I was tasked on deploying a model made externally on high resolution image to find palm trees.

    Prikaži ovu nit
    Poništi
  9. 21. sij

    This isn't perfect, far from it. We can see some misses as well as some incorrectly labelled trees. But to know the shade management of a field, this is pretty damn good, especially using free image.

    Prikaži ovu nit
    Poništi
  10. 21. sij

    The result is shown here in transparent above the higher resolution image. White pixels have been classified as shade trees. Keep in mind this is done ONLY with 10m resolution imagery, NOT the image used as reference.

    Prikaži ovu nit
    Poništi
  11. 21. sij

    Here is what the input Sentinel 2 (10m resolution remember) image might look like (1st image). 2nd image is a much higher resolution image of the same area, so you can actually see where the shade trees are.

    Prikaži ovu nit
    Poništi
  12. 21. sij

    In short, I give a satellite image, the location of some trees I know as well as some fields who's shade management I know (this is called the training data in machine learning); and I get a nice (hopefully) map of all the areas where shade trees are supposed to be.

    Prikaži ovu nit
    Poništi
  13. 21. sij

    I'm gonna detail what I came up with in the end, Twitter isn't suited for this. But for those who might care, and just because I think this looks pretty cool (and I'm proud of it), this is what my last model looks like. (Feel free to reach out / @ me if you have any questions!)

    Prikaži ovu nit
    Poništi
  14. 21. sij

    Okay, obviously this is trickier than I expected. But it's also really interesting! Most computer vision problems tackle objects that take up a large area in the image, rarely 1 or even less than 1 pixel. So there's a lot of interesting and novel things here!

    Prikaži ovu nit
    Poništi
  15. 21. sij

    Then I can threshold this (every proba > threshold becomes a tree, everything < becomes not-a-tree). This looks nice, but remember we're looking for scattered trees, and here we are seeing big patches. That's still not satisfying. (For those who care, this gets a 0.3 F1 score)

    Prikaži ovu nit
    Poništi
  16. 21. sij

    I then feed this into a model that is going to look at every pixel and give it a probability based on how close it looks to all of these pixels I marked in orange previously. Lower left there's a city (masked out) with low proba around it. Makes sense: no trees around the city.

    Prikaži ovu nit
    Poništi
  17. 21. sij

    The next idea, was trying to make a model find those patterns by itself. So I took my sharpest digital pen (aka a mouse) and started marking all the places where I, the human, could see shade trees. (Original image (cloud free Sentinel 2 + Pleiades under) and marked image)

    Prikaži ovu nit
    Poništi
  18. 21. sij

    Getting there, 61.5% accuracy. But let's be real, that's not usable. Nobody wants something that is right 6 times out of 10, or more importantly wrong 4 times out of 10. So I move on.

    Prikaži ovu nit
    Poništi
  19. 21. sij

    But that was only with 1 image! So my little brain goes on thinking and maybe trees have a different property over a year than the rest of the field? I decide to use all the cloud free images I have for 2018 and try again:

    Prikaži ovu nit
    Poništi
  20. 21. sij

    The most important number is the one in the red box: 55.5% accuracy. That means, if you give me a random crop of image, something like the picture here, my model can tell you 55 times out of 100 correctly if it's a shade tree or not. At random, accuracy is 50%. Not great.

    Prikaži ovu nit
    Poništi
  21. 21. sij

    In short the answer is no. I tried a bunch of stuff (GLCM study for those who care) and I plugged those in a machine learning algorithm (funny to use a Random Forest when looking for trees by the way), and these are the results:

    Prikaži ovu nit
    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·