Regional Attention Network (RAN) for Head Pose and Fine-Grained Gesture Recognition

Affect is often expressed via non-verbal body language such as actions/gestures, which are vital indicators for human behaviors. Recent studies on recognition of fine-grained actions/gestures in monocular images have mainly focused modeling spatial configuration parts representing pose, human-objects interactions and variations local appearance. The results show that this a brittle approach since it relies accurate parts/objects detection. In work, we argue there exist discriminative semantic regions, whose “informativeness” can be evaluated by the attention mechanism inferring gestures/actions. To end, propose novel end-to-end regional network (RAN) , fully convolutional neural (CNN) to combine multiple contextual regions through mechanism, focusing most relevant given task. Our consist one or more consecutive cells adapted from strategies used computing HOG (Histogram Oriented Gradient) descriptor. model extensively ten datasets belonging 3 different scenarios: 1) head pose recognition, 2) drivers state 3) action facial expression recognition. proposed outperforms state-of-the-art considerable margin metrics.