high performance computing on graphics processing units: hgpu.org

Random forests are a machine learning method that has recently become popular in the computer vision community to solve image segmentation and object detection tasks. Existing random forest implementations are either general purpose and not efficiently applicable for image segmentation or focus only on the speed of prediction. The implementation for the Microsoft Kinect gaming platform, for instance, achieves real-time speed on a single Microsoft Xbox GPU to recognize the pose of the user. Random forest training, however, has been conducted on a large cluster with 1000 CPU cores. Generally, training on large datasets is computationally demanding and impedes scientific research since the process takes long if a computing cluster is not available or too expensive for the task at hand. It is the goal of this master’s thesis to accelerate training and prediction of random forests for object-class image segmentation on RGB-D datasets by efficiently using CPUs and the massively parallel computing power offered by GPUs. We present an implementation that runs up to 28 times faster on GPU and is capable to train a random forest in less than four minutes on a GPU; thus drastically abbreviating a process that previously took about one whole day on a CPU. Dense classification of RGB-D images in VGA resolution runs in real-time speed on a single mobile GPU.