5 ms·
I was looking for the deep learning. But there is none in the patent and I don't think there is any in the process. "Method 100 automatically processes FMV dat
by debbiedowner 7y ago
I was looking for the deep learning. But there is none in the patent and I don't think there is any in the process.
"Method 100 automatically processes FMV data for delivering rapid two-dimensional (2-D) and three-dimensional (3-D) geospatial outputs." According to [0041] 'Method 100' is FIG 1? Looking at FIG 1, it talks a lot about clusters and metadata. From [0026] and [0039] and [0064] it seems like a cluster might be a collection of images containing a point of interest. [0089] says they do measure the altitude of the aircraft. Since they have the metadata of camera FOV they can place each image in a 3D world. I think they take a lot of redundant images and they have many views of each object of interest, which is enough to give them a (x,y,z) estimate for their points of interest.
"The increased processing speed of the system is achieved by configuration of system components to perform novel geometric calculations rather than by increased processing capability."
This gives me the idea that there is a ridiculous amount of data per scene.
I suppose when you control your own data collection, the methods can be completely different than what is worked on in academia. Compare with single image depth estimates, or even for depth maps from video for self driving.
I guess it's good that deep learning isn't being used here. Now that I think of it, it is not appropriate.
I like this part lol: "Method 100 further may be automated
using various programming scripts and tools such as a novel
Python script generator."
- debbiedowner 7y agoI pursued this a little further out of curiosity and found the inventor's CV: [0] He had a publication on there belonging to "9th NATO Military Sensing Symposium, 2017" from which papers/presentations are made public: [1], surprising! But I didn't find Massaro's paper there. I also searched for his "Videogrammetric Mapping from Unmanned Aerial Systems" but did not find it in ISPRS or elsewhere. A commercially available software that seems to use a similar method: [2] The Techcrunch article is likely a derivative of the earlier Techlink article: [3] [0]: http://mason.gmu.edu/~rmassaro/CV_Massaro_complete_20180816.pdf http://mason.gmu.edu/~rmassaro/CV_Massaro_complete_20180816.... [1]: https://www.sto.nato.int/publications/STO%20Meeting%20Proceedings/Forms/Meeting%20Proceedings%20Document%20Set/docsethomepage.aspx?ID=43014&FolderCTID=0x0120D5200078F9E87043356C409A0D30823AFA16F602008CF184CAB7588E468F5E9FA364E05BA5&List=7e2cc123-6186-4c30-8082-1ba072228ca7&RootFolder=%2Fpublications%2FSTO%20Meeting%20Proceedings%2FSTO%2DMP%2DSET%2D241 https://www.sto.nato.int/publications/STO%20Meeting%20Procee... [2]: https://www.pix4d.com/blog/underwater-mapping-videogrammetry https://www.pix4d.com/blog/underwater-mapping-videogrammetry [3]: https://techlinkcenter.org/new-us-army-software-rapidly-converts-live-drone-video-into-2d-and-3d-maps/ https://techlinkcenter.org/new-us-army-software-rapidly-conv...
- thebooktocome 7y agoIME, introducing deep learning to a problem like this causes more problems than it solves.
- varjag 7y agoYup, there is no need to involve deep learning if you have an efficient algorithmic solution pinned down. Too often people try to use AI as a substitute for thought on the normal, tractable problems.
- SiempreViernes 7y agoThe point is to reach the singularity, so by making computers smarter and people not think, you advance this project two steps for the price of one! ;)
- namibj 7y agoI'll throw https://github.com/AIBluefisher/EGSfM https://github.com/AIBluefisher/EGSfM into the discussion as an example for an AFAIK state-of-the-art structure from motion tool. It takes images/video, and reconstructs the relative positioning of the individual images/frames to each other, as well as the camera parameters required for multi-view stereo techniques like https://arxiv.org/abs/1903.10929 https://arxiv.org/abs/1903.10929 to generate meshes.