Attributes and Action Recognition Based on Convolutional Neural Networks and Spatial Pyramid VLAD Encoding

Abstract

Determination of human attributes and recognition of actions in still images are two related and challenging tasks in computer vision, which often appear in fine-grained domains where the distinctions between the different categories are very small. Deep Convolutional Neural Network (CNN) models have demonstrated their remarkable representational learning capability through various examples. However, the successes are very limited for attributes and action recognition as the potential of CNNs to acquire both of the global and local information of an image remains largely unexplored. This paper proposes to tackle the problem with an encoding of a spatial pyramid Vector of Locally Aggregated Descriptors (VLAD) on top of CNN features. With region proposals generated by Edgeboxes, a compact and efficient representation of an image is thus produced for subsequent prediction of attributes and classification of actions. The proposed scheme is validated with competitive results on two benchmark datasets: 90.4% mean Average Precision (mAP) on the Berkeley Attributes of People dataset and 88.5% mAP on the Stanford 40 action dataset.