Interactive volume classification and visualisation using incremenal-svm and cuda
Authors: David M. Hughes, Ik Soo Lim, Nigel John, Keith Hughes and Tom Rippeth
This paper presents a work-in-progress interactive tool for the visualization of volumetric data using higher-dimensional classification. The conversion of volume cells to multi-dimensional training examples enables classification results not achievable by standard DVR transfer-function methods. A user specifies input for a Support Vector Machine (SVM) which is trained to predict the classification of each remaining voxel in a volume. By utilizing Incremental SVM, our segmentation tool is able to hide the training process during the user painting period such that any delay between the moment the user stops painting to the final visualization is minimized. In order to accelerate the class prediction of the entire volume we employ CUDA-enabled GPUs, which results in more than a ten-fold decrease in the time needed when
compared to CPU-based methods.