SPORE: Staged Probabilistic Regression for Hand Orientation Inference

Learning the global hand orientation from 2D monocular images is a challenging task, as the projected hand shape is affected by a number of variations. These include inter-person hand shape and size variations, intra-person pose and style variations and self-occlusion due to varying hand orientation. Given a hand orientation dataset containing these variations, a single regressor proves to be limited for learning the mapping of hand silhouette images onto the orientation angles. We address this by proposing a staged probabilistic regressor (SPORE) which consists of multiple expert regressors, each one learning a subset of variations from the dataset. Inspired by Boosting, the novelty of our method comes from the staged probabilistic learning, where each stage consists of training and adding an expert regressor to the intermediate ensemble of expert regressors. Unlike Boosting, we marginalize the posterior prediction probabilities from each expert regressor by learning a marginalization weights regressor, where the weights are extracted during training using a Kullback-Leibler divergence-based optimization. We extend and evaluate our proposed framework for inferring hand orientation and pose simultaneously. In comparison to the state-of-the-art of hand orientation inference, multi-layered Random Forest marginalization and Boosting, our proposed method proves to be more accurate. Moreover, experimental results reveal that simultaneously learning hand orientation and pose from 2D monocular images significantly improves the pose classification performance.