Scene flow describes the motion of 3D objects in real world and potentiallycould be the basis of a good feature for 3D action recognition. However, itsuse for action recognition, especially in the context of convolutional neuralnetworks (ConvNets), has not been previously studied. In this paper, we proposethe extraction and use of scene flow for action recognition from RGB-D data.Previous works have considered the depth and RGB modalities as separatechannels and extract features for later fusion. We take a different approachand consider the modalities as one entity, thus allowing feature extraction foraction recognition at the beginning. Two key questions about the use of sceneflow for action recognition are addressed: how to organize the scene flowvectors and how to represent the long term dynamics of videos based on sceneflow. In order to calculate the scene flow correctly on the available datasets,we propose an effective self-calibration method to align the RGB and depth dataspatially without knowledge of the camera parameters. Based on the scene flowvectors, we propose a new representation, namely, Scene Flow to Action Map(SFAM), that describes several long term spatio-temporal dynamics for actionrecognition. We adopt a channel transform kernel to transform the scene flowvectors to an optimal color space analogous to RGB. This transformation takesbetter advantage of the trained ConvNets models over ImageNet. Experimentalresults indicate that this new representation can surpass the performance ofstate-of-the-art methods on two large public datasets.
展开▼
