ALBERT

Description: In this paper, we propose a novel and efficient framework for 3D action recognition using a deep learning architecture. First, we develop a 3D normalized pose space that consists of only 3D normalized poses, which are generated by discarding translation and orientation information. From these poses, we extract joint features and employ them further in a Deep Neural Network (DNN) in order to learn the action model. The architecture of our DNN consists of two hidden layers with the sigmoid activation function and an output layer with the softmax function. Furthermore, we propose a keyframe extraction methodology through which, from a motion sequence of 3D frames, we efficiently extract the keyframes that contribute substantially to the performance of the action. In this way, we eliminate redundant frames and reduce the length of the motion. More precisely, we ultimately summarize the motion sequence, while preserving the original motion semantics. We only consider the remaining essential informative frames in the process of action recognition, and the proposed pipeline is sufficiently fast and robust as a result. Finally, we evaluate our proposed framework intensively on publicly available benchmark Motion Capture (MoCap) datasets, namely HDM05 and CMU. From our experiments, we reveal that our proposed scheme significantly outperforms other state-of-the-art approaches.

Description: We propose an efficient and novel architecture for 3D articulated human pose retrieval and reconstruction from 2D landmarks extracted from a 2D synthetic image, an annotated 2D image, an in-the-wild real RGB image or even a hand-drawn sketch. Given 2D joint positions in a single image, we devise a data-driven framework to infer the corresponding 3D human pose. To this end, we first normalize 3D human poses from Motion Capture (MoCap) dataset by eliminating translation, orientation, and the skeleton size discrepancies from the poses and then build a knowledge-base by projecting a subset of joints of the normalized 3D poses onto 2D image-planes by fully exploiting a variety of virtual cameras. With this approach, we not only transform 3D pose space to the normalized 2D pose space but also resolve the 2D-3D cross-domain retrieval task efficiently. The proposed architecture searches for poses from a MoCap dataset that are near to a given 2D query pose in a definite feature space made up of specific joint sets. These retrieved poses are then used to construct a weak perspective camera and a final 3D posture under the camera model that minimizes the reconstruction error. To estimate unknown camera parameters, we introduce a nonlinear, two-fold method. We exploit the retrieved similar poses and the viewing directions at which the MoCap dataset was sampled to minimize the projection error. Finally, we evaluate our approach thoroughly on a large number of heterogeneous 2D examples generated synthetically, 2D images with ground-truth, a variety of real in-the-wild internet images, and a proof of concept using 2D hand-drawn sketches of human poses. We conduct a pool of experiments to perform a quantitative study on PARSE dataset. We also show that the proposed system yields competitive, convincing results in comparison to other state-of-the-art methods.