Patrick Marais  of the Computer Science department will present the School of IT Colloquium with a talk entitled "Semi-automated cleaning of heritage range scans with machine learning".

Abstract:

Range scans acquired by laser scanners are an important source of data in attempts to capture and preserve cultural heritage sites. Unfortunately this data almost always contains unwanted sample points, arising from people, trees and cars as well as scanner artefacts. Removing  this data -  'cleaning' the scans -  is generally a time consuming and  labour intensive process, particularly for scanning campaigns, which may involve hundreds of scans each consisting of millions of points.

In this talk I present recent work, with colleagues in Italy, on “semi-automated”  range scan cleaning. The approach recasts point cloud cleaning as a simple binary classification problem, and develops a user-driven approach to build a classifier that improves over time. The user corrects misclassifications from the predictor and these are then used to steer it towards better classification of later scans in the sequence. Preliminary tests show that the amortized processing time cost when including training, prediction and manual re-labelling, is significantly better than the time required for a human expert to label/classify all the scans in a campaign by hand.

Bio:

Patrick Marais is an associate professor in the Department of Computer Science at the University of Cape Town. His interests include computer vision applications, particularly in the cultural heritage domain, procedural graphics for content generation and scalable GPU algorithms for very large data sets. His recent work has focused mainly on the application of machine learning to point cloud processing tasks in the cultural heritage domain.