Third progress presentation

Vegard Bergsvik Øvstegård

Fri - 13 Nov 2020



Days until delivery: 186 days

GIL-UAV

GPS-Independent Localization for UAVs

System diagram
System diagram

Framework simulation

Status from previous progress presentation

Task Progress
In progress:
1. Tune the U-net**
2. Acquire & improve Dataset*
3. Train the U-net*
To do:
4. Code dataset-producing software
5. Get drone footage
6. Implement framework(C++, SIMD, CUDA)
Completed
Create U-net (PyTorch, multi-GPU)*

Updated tasks* New tasks**

Updated tasks

Task Progress
2. Acquire & improve Dataset
3. Train the U-net
4. Code dataset-producing software

Acquire & improve Dataset

  • Orthophotos:
    • Location: Norway
    • Year taken: 2018-2020
    • Ca 64 map-tiles.
    • Spanning over 200$ km^2$
  • Classes:
    • Buildings
    • Water bodies
    • Background

DS Version 0.0.1:

  • 25,1Gb of data
  • 67219 samples
  • Issues:
    • Poor building annotations due to rectification
    • Missing annotations on some buildings
    • Very inaccurate water annotations

DS Bump to Version 0.1.0:

  • Improvements:
    • Removed sets with images out of bounds from aerial photos.
    • Removed sets containing only background class.
    • Removed sets containing only water bodies class.
  • 17,4Gb of data
  • 43802 samples
  • Class balance:
    • Building percentage: 24 %
    • Water pixels: 6 %
    • Background pixels: 69 %

Code dataset-producing software

Features:

  • Automated extraction of ground truth labels from vector maps.
  • Generates datasets from orthophotos and GT images.
  • Discards (some)noisy sets.
  • Produces dataset statistics
    • Class balance
    • Mean
    • Standard deviation

Backlog:

  • Remove sets with non-annotated buildings.

Train the U-net

Test set example from newest dataset:

Images and predictions
Images and predictions

Train the U-net

Ground truth and predictions
Ground truth and predictions

Current status and progress

Task Progress
In progress:
1. Tune the U-net
2. Acquire & improve Dataset*
3. Train the U-net*
To do:
4. Get drone footage
5. Implement framework(C++, SIMD, CUDA)
Completed
Code dataset-producing software

Updated tasks* New tasks**

Completed tasks

  • Code dataset-producing software
  • Create U-net (PyTorch, multi-GPU)
  • Implement naive MCL algorithm (Python)
  • Get hardware (nVIDIA Jetson TX1)

Plan for the next fortnight:



Week 47 Week 48
Tune and train the U-net Port MCL code from Python to C++
Get drone footage Get drone footage
Acquire & improve Dataset Create orthophotos from drone footage

Print

Back