Robosat: an Open Source and efficient Semantic Segmentation Toolbox - - PowerPoint PPT Presentation

▶
robosat an open source and efficient semantic
SMART_READER_LITE
LIVE PREVIEW

Robosat: an Open Source and efficient Semantic Segmentation Toolbox - - PowerPoint PPT Presentation

Mar 05, 2024 31 likes •393 views

Robosat: an Open Source and efficient Semantic Segmentation Toolbox for Aerial Imagery @o_courtin @PyParisFr 2018 RoboSat Generic ecosystem for QoD and feature extraction from aerial and satellite imagery https://github.com/mapbox/robosat

slide-1
SLIDE 1

Robosat: an Open Source and efficient Semantic Segmentation Toolbox for Aerial Imagery

@o_courtin @PyParisFr 2018

slide-2
SLIDE 2

https://github.com/datapink/robosat https://github.com/mapbox/robosat

RoboSat

Generic ecosystem for QoD and feature extraction from aerial and satellite imagery

slide-3
SLIDE 3

RoboSat

State of Art SemSeg Industrial standards code design and written Higly modular and quite extensible OSM and MapBox ecosystem integration PyTorch based Licence MIT

slide-4
SLIDE 4

Neurons Network Data Labels Weights Loss Function

slide-5
SLIDE 5

Neurons Network Data Labels Weights Loss Function

slide-6
SLIDE 6

Download WMS TMS XYZ Rasterize GeoJSON Extract OSM pbf Cover Image Tile Raster Label Subset Training DataSet Bbox XYZ dir

Data Preparation

slide-7
SLIDE 7

Neurons Network Data Labels Weights Loss Function

slide-8
SLIDE 8

https://arxiv.org/pdf/1806.00844.pdf

slide-9
SLIDE 9

MultiBands and Fusion

Multi spectral imagery

  • r any (related) vector rasterization
slide-10
SLIDE 10

Neurons Network Data Labels Weights Loss Function

slide-11
SLIDE 11

Image Label Cross Entropy mIoU Lovasz http://www.cs.toronto.edu/~wenjie/papers/iccv17/mattyus_etal_iccv17.pdf http://www.cs.umanitoba.ca/~ywang/papers/isvc16.pdf https://arxiv.org/abs/1705.08790

slide-12
SLIDE 12

Neurons Network Data Labels Weights Loss Function

slide-13
SLIDE 13

Weights

ImageNet pre-trained Resume Training Export ONNX

slide-14
SLIDE 14

Grand Lyon OpenData use case

slide-15
SLIDE 15

Rasterize Images Lyon GeoJson Labels Download Lyon WMS

slide-16
SLIDE 16

Rasterize Images Lyon GeoJson Labels Train Labels Val Labels Train Images Val Images Subset Subset Download Lyon WMS

slide-17
SLIDE 17

Rasterize Images Lyon GeoJson Labels Train Labels Val Labels Train Images Val Images Subset Subset Train Model Download Lyon WMS

slide-18
SLIDE 18

Rasterize Images Lyon GeoJson Labels Train Labels Val Labels Train Images Val Images Subset Subset Train Model Predict Masks Download Lyon WMS Images

slide-19
SLIDE 19

Rasterize Images Lyon GeoJson Labels Train Labels Val Labels Train Images Val Images Subset Subset Train Model Predict Masks Download Lyon WMS Images Compare

slide-20
SLIDE 20
slide-21
SLIDE 21
slide-22
SLIDE 22
slide-23
SLIDE 23

Rasterize Images Lyon GeoJson Labels Train Labels Val Labels Train Images Val Images Subset Subset Train Model Predict Masks Download Lyon WMS Images Compare

slide-24
SLIDE 24

Rasterize Images Lyon GeoJson Labels Train Labels Val Labels Train Images Val Images Subset Subset Train Model Predict Masks OSM GeoJson Rasterize OSM Masks Compare Download Lyon WMS Images Compare

slide-25
SLIDE 25
slide-26
SLIDE 26
slide-27
SLIDE 27

Stacks

Proj 4 GEOS GDAL Rasterio Shapelib CUDA cuDNN PyTorch NumPy OpenCV RoboSat

slide-28
SLIDE 28

August v 1.0.0 Initial release daniel-j-h bkowshik September v 1.1.0 Increase Training perfs Jesse-jApps ocourtin October master OSM Roads extraction DragonEmperorG mIoU and Lovasz losses

  • courtin

November PR 138 Multibands and tools stuff

  • courtin

Code reviewer since ever : daniel-j-h :)

RoboSat Timeline

slide-29
SLIDE 29

Next ?

#1 Predict performance improvment #2 Lower resolution Imagery SemSeg: Sentinel-2 or PlanetLab #3 Feature extraction

slide-30
SLIDE 30

Next ?

#1 Predict performance improvment

  • PyTorch 1.0 JIT
  • CUDA 10 FP 16 models
  • ONNX export to high performance env (Caffe2 / Microsoft ?)
  • Lighter models
slide-31
SLIDE 31

Next ?

#2 Lower resolution Imagery SemSeg: Sentinel-2 or PlanetLab

  • Improve again Fusion and Topological Losses
  • SuperPixel resolution
slide-32
SLIDE 32

Next ?

#3 Feature extraction

  • Generic feature post treatment. Explore GAN
slide-33
SLIDE 33

Next ?

Predict performance improvment Lower resolution Imagery SemSeg: Sentinel-2 or PlanetLab Feature extraction

slide-34
SLIDE 34

Take Away

Industrial state of art Aerial SemSeg available, and playful Data are also available Decent OpenDataSet is a bottle neck Predict speed performances had to been improve to scale at large

slide-35
SLIDE 35

Computer Vision NLP TimeSeries www.datapink.com @data_pink expertise, core dev and solutions :

slide-36
SLIDE 36

Computer Vision NLP TimeSeries www.datapink.com @data_pink expertise, core dev and solutions : Coming conf, 05/12 @OSS_Paris : NLP State of Art