Multi-layer random forests with auto-context for object detection and pose estimation

Abstract

Deep neural networks are multi-layer architectures of neural networks used in machine learning. In recent years, training such systems through Deep Learning techniques has shown remarkable improvements over more traditional techniques in the domains of image-based classification and object category recognition. However, a well-known downside of Deep Learning is that generally, for training these very-high-dimensional systems, a very large amount of training samples is required. It is a common understanding that the multi-level structure of data processing that is learned and applied may be a critical factor for the success story of deep neural networks. Highly task-optimized representations may be gradually built through a sequence of transformations that can abstract and hence generalize from the specific data instances. In this Master Thesis, the aspect of deep multi-layered data transformation will be transferred from neural networks onto a very different learning scheme, the Random Forests. The latter is a highly competitive and general-purpose method for classification and regression. Even large Random Forests usually do not have the high amount of parameters to optimize during training, and the number of hyper-parameters and architectural varieties is much lower than for the deep neural networks. Training with a much smaller amount of samples is hence possible. Some variants of a layered architecture of Random Forests is investigated here, and different styles of training each forest is tried. The specific problem domain considered here is object detection and pose estimation from RGB-D images. Hence, the forest output is used by the pose hypothesis scoring function and the poses are optimized using RANSAC-based scheme. Experiments on the pose annotated Hinterstoisser and T-LESS datasets prove the performance of the multi-layer random forests architecture.