This report describes experiments conducted using the multi-class image classification framework implemented in the stair vision library (SVL, (Gould et al., 2008)) in the context of the ISPRS 2D semantic labeling benchmark. The motivation was to get results from a well-established and public available software (Gould, 2014), as a kind of baseline. Besides the use of features implemented in the SVL which makes use of three channel images, assuming RGB, we also included features derived from the height model and the NDVI which is specific here, because the benchmark dataset provides surface models and CIR images. Another point of interest concerned the impact the segmentation had on the overall result. To this end a pre-study was performed where different parameters for the graph-based segmentation method introduced by Felzenszwalb and Huttelocher (2004) have been tested, in addition we only applied a simple chessboard segmentation. Other experiments focused on the question whether the conditional random field classification approach helps to enhance the overall performance. The official evaluation of all experiments described here is available at http://www2.isprs.org/vaihingen-2d-semantic-labeling-contest.html (SVL_1 to SVL_6). The normalized height models are available through the ReseachGate profile of the author (http://www.researchgate.net/profile/Markus_Gerke)
[1]
Jaewook Jung,et al.
Results of the ISPRS benchmark on urban object detection and 3D building reconstruction
,
2014
.
[2]
P. Axelsson.
DEM Generation from Laser Scanner Data Using Adaptive TIN Models
,
2000
.
[3]
智一 吉田,et al.
Efficient Graph-Based Image Segmentationを用いた圃場図自動作成手法の検討
,
2014
.
[4]
E. Baltsavias,et al.
A TEST OF AUTOMATIC ROAD EXTRACTION APPROACHES
,
2006
.
[5]
GouldStephen,et al.
Multi-Class Segmentation with Relative Location Prior
,
2008
.
[6]
Stephen Gould,et al.
Multi-Class Segmentation with Relative Location Prior
,
2008,
International Journal of Computer Vision.