1 / 1

3D Laser Scan Classification Using Web Data and Domain Adaptation Kevin Lai, Dieter Fox {kevinlai, fox}@cs.washington.e

3D Laser Scan Classification Using Web Data and Domain Adaptation Kevin Lai, Dieter Fox {kevinlai, fox}@cs.washington.edu. Department of Computer Science & Engineering University of Washington, Seattle. Objective

reuel
Download Presentation

3D Laser Scan Classification Using Web Data and Domain Adaptation Kevin Lai, Dieter Fox {kevinlai, fox}@cs.washington.e

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. 3D Laser Scan Classification Using Web Data and Domain Adaptation Kevin Lai, Dieter Fox{kevinlai, fox}@cs.washington.edu Department of Computer Science & EngineeringUniversity of Washington, Seattle Objective - Significantly reduce the need for manually labelled training data by leveraging data sets available on the World Wide Web. - Use models from Google 3D Warehouse to train classifiers for 3D laser scans collected by a robot navigating through urban environments and perform domain adaptation to deal with the different characteristics of the two data sets. System Overview - A separate distance function (linear combination of feature differences) for each training exemplar. - Exemplars within a class (e.g. car) can be very different from each other, so the system should identify subsets of them that are similar. - Learn these via an iterative optimization process that uses an SVM. - Shape features: Spin Image Signatures, bounding box dimensions. - We investigate two techniques for Domain Adaptation: 1. Feature Augmentation (Daumé 2007), referred to as Stacked in the experiments section. 2. A technique specific to exemplar-based distance learning introduced by us, referred to as Alpha in the experiments section. - Model the reliability of the training exemplars. - Perform segmentation on a 3D laser scan and obtain a probability distribution over the class of each laser point. Example Matches The leftmost column shows example segments extracted from 3D laser scans at test time: car, person, tree (top to bottom). Second to last columns show training exemplars with distance below threshold, closer exemplars are further to the left. Example Classification Experiments Precision-recall curves comparing: (Left) Performance of the various approaches, trained five laser scans. (Right) Performance of the various approaches, trained on three laser scans. (Top) Ground truth classification for part of a 3D laser scan. (Bottom) Classification achieved by our approach. Colors indicate ground plane (cyan) and object types (green: tree, blue: car, yellow: street sign, purple: person, red: building, grey: other, white: not classified). Precision-recall curves comparing the performance of our probabilistic classification technique and the recognition confidence scoring technique of Malisiewicz and Efros: (Left) Trained on five laser scans. (Right) Trained on three laser scans. Acknowledgements: This work was supported in part by a ONR MURI grant number 5710002230, by the NSF under contract number IIS-0812671, and by a postgraduate scholarship from the Natural Sciences and Engineering Research Council of Canada. We would also like to thank Michael Beetz for the initial idea to use 3D-Warehouse data for object classification.

More Related