© Kaaenaat Private Limited1
Akbar Ladak
ladak@kaaenaat.com
Object Classification in 3D:
Working with LiDAR point clouds
© Kaaenaat Private Limited2
© Kaaenaat Private Limited3
© Kaaenaat Private Limited4
Challenge adapting to PointClouds:Scale
Images
12.4 MB
• Uncompressed RGB 16-bit frame, 2.1 MP, 24 bit
•
Point Clouds
~85 MB
• Point Cloud LAS file with 2.1 m points
• Most files between 250 to 500 MB
© Kaaenaat Private Limited5
Challenge adapting to PointClouds:Tools
Images
OpenCV
• Mature toolset with
time-tested algorithms
• High fidelity cameras with
relatively low noise
Point Clouds
PCL
• Theoretical Underpinnings unable to adapt
to real world, noisy data
• Needed to create techniques
that would be robust to noisy data
© Kaaenaat Private Limited6
OurWork (More details to follow)
1. Creating 3D pattern recognition techniques robust to noise
2. Dimensionality reduction for Machine Learning using 3D
trigonometry, calculus & heuristics
3. Generative Adversarial Networks to create new diverse
datasets to extend the applicability of the algorithm

Anthill inside-2017-talk-proposal

  • 1.
    © Kaaenaat PrivateLimited1 Akbar Ladak ladak@kaaenaat.com Object Classification in 3D: Working with LiDAR point clouds
  • 2.
  • 3.
  • 4.
    © Kaaenaat PrivateLimited4 Challenge adapting to PointClouds:Scale Images 12.4 MB • Uncompressed RGB 16-bit frame, 2.1 MP, 24 bit • Point Clouds ~85 MB • Point Cloud LAS file with 2.1 m points • Most files between 250 to 500 MB
  • 5.
    © Kaaenaat PrivateLimited5 Challenge adapting to PointClouds:Tools Images OpenCV • Mature toolset with time-tested algorithms • High fidelity cameras with relatively low noise Point Clouds PCL • Theoretical Underpinnings unable to adapt to real world, noisy data • Needed to create techniques that would be robust to noisy data
  • 6.
    © Kaaenaat PrivateLimited6 OurWork (More details to follow) 1. Creating 3D pattern recognition techniques robust to noise 2. Dimensionality reduction for Machine Learning using 3D trigonometry, calculus & heuristics 3. Generative Adversarial Networks to create new diverse datasets to extend the applicability of the algorithm