Action recognition in static image is challenging. The authors propose mutually incoherent pose bases which are implicit poselet co‐occurrences and are learned by dictionary training to describe body pose. Poselets in a pose basis are not constrained in space and quantity, thus pose basis can describe body pose more flexibly than k ‐poselet. In their method, body pose in an image is represented by a sparse linear combination of pose bases because pose in an action varies while each image only captures a snapshot from a single viewpoint. In dictionary training, the challenge is how to stabilise the sparse representation which is the input of Support Vector Machine (SVM) for action recognition, because the original pose signal is ambiguous while dictionary is an over complete matrix. Their solution is to add cumulative coherence as penalty in objective function and induce pose bases become mutually incoherent. They evaluate the method on two popular datasets and experiment results show the pose representation has encouraging performance in action recognition. Furthermore, they empirically exploit the complementary role of the local pose feature with deep convolutional neural network features from holistic image. Experiment results demonstrate aggressive performance improvement by concatenating the two features.

We propose mutually incoherent pose bases for action recognition in static image, each of which implicitly represents co-occurrence of poselets. First of all, action specific poselets are trained. To suppress the ambiguity of detection, we cluster poselet activations by the overlap of predicted torso bound of each poselet. Then pose feature of an action person can be extracted which is a vector composed of poselet detection. In dictionary training, our challenge is that dictionary is over complete thus small perturbation in pose feature would cause significant change in sparse code, which might change classification result. In our framework, a penalty which induces pose bases become mutually incoherent is added to the objective function. We evaluate the method on PASCAL VOC 2012 Action dataset and Ikizler 5-Action dataset, experiment results show wonderful performance compared with counterparts and baselines.

This paper addresses the problem of action recognition from body pose. Detecting body pose in static image faces great challenges because of pose variability. Our method is based on action-specific hierarchical poselet. We use hierarchical body parts each of which is represented by a set of poselets to demonstrate the pose variability of the body part. Pose signature of a body part is represented by a vector of detection responses of all poselets for the part. In order to suppress detection error and ambiguity we explore to use part-based model (PBM) as detection context. We propose a constrained optimization algorithm for detecting all poselets of each part in context of PBM, which recover neglected pose clue by global optimization. We use a PBM with hierarchical part structure, where body parts have varying granularity from whole body steadily decreasing to limb parts. From the structure we get models with different depth to study saliency of different body parts in action recognition. Pose signature of an action image is composed of pose signature of all the body parts in the PBM, which provides rich discriminate information for our task. We evaluate our algorithm on two datasets. Compared with counterpart methods, pose signature has obvious performance improvement on static image dataset. While using the model trained from static image dataset to label detected action person on video dataset, pose signature achieves state-of-the-art performance.

General object detection still remains a big challenge for vision researchers. In this paper, we are particularly interested in the subject of object detection in the context of street scene. Our image database consists of video frames taken from urban street which tends to be crowded and presents a lot of artificial objects. Traditional street scene understanding methods often involve 3D reconstruction of the street scene before object detection. We argue that through carefully-chosen features and utilizing category-dependent detectors, we can still achieve good detection performance thus gain good understanding of street scene by merely low quality 2D images. In our detection framework,we use hybrid detectors for different object categories. For example, basic SVM classifier is adopted to detect rigid objects like traffic lights, traffic sign, lamp and fire hydrant; texture objects like trees are detected via a discriminative texture classifier; while for semi-rigid and multiple view objects like cars, votingbased detector is applied. We further prune false positives by utilizing appearance cues. Experiment result shows our method is able to recognize meaningful objects on street and gives attention to drivers or directions to auto-driven vehicles.

Manifold clustering aims to partition a set of input data into several clusters each of which contains data points from a separate, simple low-dimensional manifold. This paper presents a novel solution to this problem. The proposed algorithm begins by randomly selecting some neighboring orders of the input data and defining an energy function that is described by geometric features of underlying manifolds. By minimizing such energy using the tabu search method, an approximately optimal sequence could be found with ease, and further different manifolds are separated by detecting some crucial points, boundaries between manifolds, along the optimal sequence. We have applied the proposed method to both synthetic data and real image data and experimental results show that the method is feasible and promising in manifold clustering.

Video Stabilization is one of those important video processing techniques to remove the unwanted camera vibration in a video sequence. In this paper, we present a practical method to remove the annoying shaky motion and reconstruct a stabilized video sequence with good visual quality. Here, the scale invariant (SIFT) features, proved to be invariant to image scale and rotation, is applied to estimate the camera motion. The unwanted vibrations are separated from the intentional camera motion with the combination of Gaussian kernel filtering and parabolic fitting. It is demonstrated that our method effectively removes the high frequency 'noise' motion, but also minimize the missing area as much as possible. To reconstruct the undefined areas, resulting from motion compensation, we adopt the mosaicing method with Dynamic Programming. The proposed method has been confirmed to be effective over a widely variety of videos.

In this paper, we propose an approach for 2D discrete vector field segmentation based on the Green function and normalized cut. The method is inspired by discrete Hodge Decomposition such that a discrete vector field can be broken down into three simpler components, namely, curl-free, divergence-free, and harmonic components. We show that the Green Function Method (GFM) can be used to approximate the curl-free and the divergence-free components to achieve our goal of the vector field segmentation. The final segmentation curves that represent the boundaries of the influence region of singularities are obtained from the optimal vector field segmentations. These curves are composed of piecewise smooth contours or streamlines. Our method is applicable to both linear and nonlinear discrete vector fields. Experiments show that the segmentations obtained using our approach essentially agree with human perceptual judgement.

Eddy tracking is just finding all the vectors with similar motion in its influence region and recording its moving route. Virtually, a flow field is able to be replaced with a scalar dataset, if its feature information can be faithfully preserved by the scalar dataset. Therefore, the eddy tracking problem can be solved by tracking the moving objects in an image sequence after a certain transformation. Here, we propose a new method to track several moving eddies simultaneously and three main technologies are introduced: green function method, motion probability distribution and quadratic program.