David Griffiths

PhD Candidate, 3D vision & machine learning

I am currently a 3rd year Ph.D candidate working on point cloud classification and detection. I am currently based in the Department of Computer Science at UCL, working under Dr. Jan Boehm and Prof. Tobias Ritschel. My research looks at developing deep learning models for understanding of 3D scenes. Specifically my Ph.D involves adressing the scarcity of manually labelled 3D data. My main academic interests include; computer vision, machine learning and photogrammetry, and in particular working at the intersection of these disiplines.

download cv

Publications.

Curiosity-driven 3D Scene Structure from Single-image Self-supervision

David Griffiths, Jan Boehm, Tobias Ritschel

arXiv preprint, 2020

We present a novel method for fully self-supervised scene paramaterisation from a single image. We achieve this by employing analysis-by-synthesis using a differentiable renderer. We show that a simple L2 loss is not sufficient for such a task. Instead, we introduce a GAN-like critic to constrain the network to propose realistic outputs. By adding such a constraint we observe the L2 loss is now sufficient to solve a number of tasks on both synthetic and real data.

PDF

     

Code

     

Project Page

Finding Your (3D) Center: 3D object detection using a learned loss

David Griffiths, Jan Boehm, Tobias Ritschel

European Conference on Computer Vision (ECCV) 2020

We present a novel approach to training for point cloud object detection. By employing two networks, we show how first a smaller local network can be trained with a fraction of the required training data. We then demonstrate how this first network can be used as a loss function to train a second full scene network without the need for labels.

PDF

     

Code

     

Project Page

SynthCity: A large scale synthetic point cloud.

David Griffiths, Jan Boehm

arXiv preprint, 2019

SynthCity is a fully synthetic Mobile Laser Scanner (MLS) point cloud. We create SynthCity using the open-source blender and blensor packages. We show that synthetic MLS datasets can be easily generated with full semantic labels. Such data can therefore be used for training machine/deep learning models.

PDF

     

Project Site

A review on deep learning techniques for 3D sensed data classification

David Griffiths, Jan Boehm

Remote Sensing, 2019

We present a full state-of-the-art review on deep learning techniques for classifcation of point cloud data. The paper covers all recent research approaches from classical machine learning methods, RGB-D, voxel, multi-view to raw point cloud processing.

PDF

     

Project Page

Improving public data for building segmentation from Convolutional Neural Networks for fused airborne lidar and image data using active contours

David Griffiths, Jan Boehm

ISPRS Journal of Photogrammetry and Remote Sensing, 2019

Manually labelling buildings for segmentation is a time consuming task. We show that readily available GIS mapping data such as that from the Ordnance Survey (UK) can be used as training data. Further, we develop a novel pipeline which uses Active Contour models and fued image-lidar data to achieve state-of-the-art accuracy in aerial building segmentation.

PDF

     

Code

     

Project Page

Weighted point cloud augmentation for neural network training data class-imbalance

David Griffiths, Jan Boehm

Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci, 2019

A key issue when training deep neural networks for outdoor point clouds is the inevitable large data imbalance. For example, a typical street scene will contain orders of magnitudes more ground points than street furniture. We develop a novel solution to apply a weighted augmentation to phyically decrease the class-imbalance.

PDF

     

Project Page

Comparison of pre-and self-calibrated camera calibration models for UAS-derived nadir imagery for a SfM application

David Griffiths, Helene Burningham

Progress in Physical Geography: Earth and Environment, 2019

Linear topologies can be challenging terrains for Structure-from-Motion pipelines. A key source of error is caused by intrinsic camera distortions. We demonstrate through effective camera pre-calibration, distortions can be significantly reduced in such applications. Ultimately increasing the accuracy of derived Digital Surface Models (DEMs).

PDF

     

Project Page

Rapid object detection systems, utilising deep learning and unmanned aerial systems (UAS) for civil engineering applications

David Griffiths, Jan Boehm

Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci, 2018

Linear topologies can be challenging terrains for Structure-from-Motion pipelines. A key source of error is caused by intrinsic camera distortions. We demonstrate through effective camera pre-calibration, distortions can be significantly reduced in such applications. Ultimately increasing the accuracy of derived Digital Surface Models (DEMs).

PDF

     

Project Page

Education.

  • 2017-Present

    Ph.D Candidate, 3D vision and machine learning, University College London

    I am currently persuing a Ph.D in 3D vision and machine learning. My primary reserarch aims at addressing data scarcity for 3D data classification and object detection. Primarily for deep learning applications.

  • 2016-2017

    MSc Remote Sensing, University College London

    Awarded an MSc in Remote Sensing with distinction. The course covered RGB and Multi-spectral image, lidar and radar data processing. My dissertation covered camera calibration for photogrammetric reconstructions in challenging topologies.

Contact.

david.griffiths.16@ucl.ac.uk
  • Dept. Computer Science,
  • 66-72 Gower Street,
  • University College London,
  • London,
  • WC1E 6EA