Griffiths, David; (2022) Label Efficient 3D Scene Understanding. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text label_efficient_3d_scene_understanding.pdf - Accepted Version Download (45MB) | Preview

Abstract

3D scene understanding models are becoming increasingly integrated into modern society. With applications ranging from autonomous driving, Augmented Real- ity, Virtual Reality, robotics and mapping, the demand for well-behaved models is rapidly increasing. A key requirement for training modern 3D models is high- quality manually labelled training data. Collecting training data is often the time and monetary bottleneck, limiting the size of datasets. As modern data-driven neu- ral networks require very large datasets to achieve good generalisation, finding al- ternative strategies to manual labelling is sought after for many industries. In this thesis, we present a comprehensive study on achieving 3D scene under- standing with fewer labels. Specifically, we evaluate 4 approaches: existing data, synthetic data, weakly-supervised and self-supervised. Existing data looks at the potential of using readily available national mapping data as coarse labels for train- ing a building segmentation model. We further introduce an energy-based active contour snake algorithm to improve label quality by utilising co-registered LiDAR data. This is attractive as whilst the models may still require manual labels, these labels already exist. Synthetic data also exploits already existing data which was not originally designed for training neural networks. We demonstrate a pipeline for generating a synthetic Mobile Laser Scanner dataset. We experimentally evalu- ate if such a synthetic dataset can be used to pre-train smaller real-world datasets, increasing the generalisation with less data. A weakly-supervised approach is presented which allows for competitive per- formance on challenging real-world benchmark 3D scene understanding datasets with up to 95% less data. We propose a novel learning approach where the loss function is learnt. Our key insight is that the loss function is a local function and therefore can be trained with less data on a simpler task. Once trained our loss function can be used to train a 3D object detector using only unlabelled scenes. Our method is both flexible and very scalable, even performing well across datasets. Finally, we propose a method which only requires a single geometric represen- tation of each object class as supervision for 3D monocular object detection. We discuss why typical L2-like losses do not work for 3D object detection when us- ing differentiable renderer-based optimisation. We show that the undesirable local- minimas that the L2-like losses fall into can be avoided with the inclusion of a Generative Adversarial Network-like loss. We achieve state-of-the-art performance on the challenging 6DoF LineMOD dataset, without any scene level labels.

Downloads since deposit

Loading...

34Downloads

Download activity - last month

Export as XMLJSONCSV

Loading...

Download activity - last 12 months

Export as XMLCSVJSON

Loading...

Downloads by country - last 12 months

Export as XMLCSVJSON

1.	United Kingdom	5
2.	United States	3
3.	Australia	1
4.	Chile	1
5.	Thailand	1
6.	Germany	1

102550all

Archive Staff Only

View Item