Video interest points, in combination with local appearance descriptors, are used for human action recognition. Most of the previously proposed video interest point detectors are straightforward extensions of some ima...
详细信息
ISBN:
(纸本)1595930361
Video interest points, in combination with local appearance descriptors, are used for human action recognition. Most of the previously proposed video interest point detectors are straightforward extensions of some image interest point detector or the other. these methods treat the temporal dimension (inter-frame) similar to the spatial dimensions (intra-frame). We argue that certain unique properties of the temporal dimension beg a different treatment. We propose an interest point detector based on vector calculus of optical flow to take advantage of the unique properties of the temporal dimension. Compared to previously proposed methods, the proposed method exhibits higher repeatability (robustness) and lower displacement (stability) of interest points under two common video transformations tested-video compression and spatial scaling. It also shows competitive action recognition performance when paired with appropriate feature descriptors in a bag of features model. Copyright is held by the authors.
We propose an example-based super-resolution (SR) framework, which uses a single input image and, unlike most of the SR methods does not need an external high resolution (HR) dataset. Our SR approach is based in spars...
详细信息
ISBN:
(纸本)1595930361
We propose an example-based super-resolution (SR) framework, which uses a single input image and, unlike most of the SR methods does not need an external high resolution (HR) dataset. Our SR approach is based in sparse representation framework, which depends on a dictionary, learned from the given test image across different scales. In addition, our sparse coding focuses on the detail information of the image patches. Furthermore, in the above process we have considered non-local combination of similar patches in the input image, which assist us to improve the quality of the SR result. We demonstrate the effectiveness of our approach for intensity images as well as range images. Contemplating the importance of edges in images of boththese modalities, we have added an edge preserving constraint that will maintain the continuity of edge related information to the input low resolution image. We investigate the performance of our approach by rigorous experimental analysis and it shows to perform better than some state-of-the-art SR approaches. Copyright 2014 ACM.
In this paper, we propose a novel method for image texture characterization. Characterization is governed by simple perceptual variations in relative orientations in terms of either no variations present or variations...
详细信息
ISBN:
(纸本)1595930361
In this paper, we propose a novel method for image texture characterization. Characterization is governed by simple perceptual variations in relative orientations in terms of either no variations present or variations present as row specific, column specific or diagonal specific. this generalization is obtained by modeling the input as a whole or image blocks depending on the broader or narrow coverage respectively. Most of the texture characterization is done either keeping a specific domain (synthetic or natural images specific to a category) or is application specific (segmentation on specific benchmark dataset or image retrieval). Contrary to this, our method is not biased towards any domain or application;rather it acts as a pre-processing step for guiding towards locating both non-textural and orientation specific textural image blocks. the proposed method quantifies the texture-tonal characterization of an image or image blocks using statistical ANOVA grading system. Once the grading for abstraction is assigned for bothimage as a whole and also for image blocks, the decision as to which higher-level algorithms need to be implemented on which block will become easier. the proposed method can be considered to be a three stage process - progressive sampling, image partitioning in blocks, ANOVA analysis and grading. Copyright 2014 ACM.
In this paper, we propose a novel image completion method using transform domain patch approximation method and kd-tree based nearest neighbor field (NNF) computation in multiscale fashion. In NNF, two important proce...
详细信息
ISBN:
(纸本)1595930361
In this paper, we propose a novel image completion method using transform domain patch approximation method and kd-tree based nearest neighbor field (NNF) computation in multiscale fashion. In NNF, two important processes are initialization of image target region and candidate patch searching method. Most of the previous techniques choose random initialization with arbitrary source image pixels or garbage values. It may misguide to image completion process and allow to select the bad candidate patches. We solve the problem using higher order singular value decomposition (HOSVD). It smoothly generates information in the target region enhancing the edge sharpness which helps to complete image structure quite successfully. It also preserves texture color in the target region. To overcome the problem of patch searching, we introduce robust kd-tree search method in our patch approximation step. Our experiment and analysis shows that the proposed method can be applied to the various types of image editing tools for natural images. Copyright 2014 ACM.
this paper proposes a novel algorithm for object tracking using Approximate Nearest Neighbour Fields (ANNF). ANNF maps have been previously used to address several problems like denoising, image completion, re-targeti...
详细信息
ISBN:
(纸本)1595930361
this paper proposes a novel algorithm for object tracking using Approximate Nearest Neighbour Fields (ANNF). ANNF maps have been previously used to address several problems like denoising, image completion, re-targeting and medical image analysis. In this paper, we deal withthe challenging problem of visual object tracking, using patch flow. the proposed method uses FeatureMatch to find patch corre- spondence between successive frames, enabling the tracker to find the best match for the object in the next frame. Based on the flow, each patch is labeled as either foreground or background. the proportion of FG/BG/border patches contributing to each pixel determines its final label. We show that objects can be successfully tracked across videos, under challenging conditions such as scale variations, illumination changes and occlusion using the proposed technique. Copyright 2014 ACM.
In this paper, we propose a framework for recognizing the group dynamics embedded in social gathering videos. We aim to identify interacting sub-groups present in a scene. In particular we are interested in identifyin...
详细信息
ISBN:
(纸本)1595930361
In this paper, we propose a framework for recognizing the group dynamics embedded in social gathering videos. We aim to identify interacting sub-groups present in a scene. In particular we are interested in identifying the participants converging to form a group, the participants dispersing from a group and the ones in static groups. We use the linear cyclic pursuit (LCP) based framework to model the collective motion. the proposed algorithm employs trajectories of individuals over a period of time to estimate the model parameters. We show that the details of group dynamics are hidden in the eigenvalues and eigenvectors of the pursuit matrix. the experiments are done on the simulated data as well as on the real gathering videos. the results show a promising scope of the proposed approach. Copyright 2014 ACM.
Action recognition from 3D depth maps/skeleton has been an active research area in computervision in the recent past. In this paper, we propose a method for action recognition by learning the sequence of various pose...
详细信息
ISBN:
(纸本)1595930361
Action recognition from 3D depth maps/skeleton has been an active research area in computervision in the recent past. In this paper, we propose a method for action recognition by learning the sequence of various poses involved while performing an action. Each pose in the sequence is represented as a set of 3D edge vectors connecting important joint points in the skeleton and 3D trajectory vectors connecting the joint point locations in the previous frame withthat in the current frame. the training samples of each action class, represented as sequences of poses, are time normalized using Dynamic Time Warping(DTW) and average pooled to construct the model sequence representing the action. Action recognition on a test sequence is achieved by finding its nearest model sequence in terms of a proposed distance measure. the effectiveness of the proposed algorithm is evaluated on two challenging datasets: Berkeley Multimodal Human Action Database and MSR Action 3D dataset. Copyright 2014 ACM.
Approaches to articulated kinematic chain analysis suffer from a lack of robustness in segmentation which affects all upstream tasks. In this work, we consider revolute chains (all joints are hinges, or ball-and-socke...
详细信息
ISBN:
(纸本)1595930361
Approaches to articulated kinematic chain analysis suffer from a lack of robustness in segmentation which affects all upstream tasks. In this work, we consider revolute chains (all joints are hinges, or ball-and-socket), and show that Sparse Spectral Clustering (SSC) is more likely to agree on tracked data points that are far from the joint. We then develop two variants of consensual clustering which marks tracks as outliers if different classifiers do not agree. this leaves a large set of points in each link, and enables robust segmentation for many real-life situations. In addition to results on Hopkins-155 [18] and Yan-Pollefeys [21, 22, 23] datasets, we also release additional datasets involving revolute chains. Copyright is held by the authors.
the proceedings contain 90 papers. the topics discussed include: a GPU based real-time CUDA implementation for obtaining visual saliency;fingerprint enhancement using unsupervised hierarchical feature learning;a Schr&...
ISBN:
(纸本)1595930361
the proceedings contain 90 papers. the topics discussed include: a GPU based real-time CUDA implementation for obtaining visual saliency;fingerprint enhancement using unsupervised hierarchical feature learning;a Schrödinger formalism for simultaneously computing the Euclidean distance transform and its gradient density;duplication detection for image sharing systems;finding group interactions in social gathering videos;particle coding for meshfree cutting of deformable assets;total cluster: a person agnostic clustering method for broadcast videos;multispectral multifocus image fusion with guided steerable frequency and improved saliency;generalized synthesis and analysis prior algorithms with application to impulse denoising;and discovery of sparse contextual relationships for activity recognition in continuous videos.
We present an approach for learning low- and high-level fingerprint structures in an unsupervised manner, which we use for enhancement of fingerprint images and estimation of orientation fields, frequency images, and ...
详细信息
ISBN:
(纸本)1595930361
We present an approach for learning low- and high-level fingerprint structures in an unsupervised manner, which we use for enhancement of fingerprint images and estimation of orientation fields, frequency images, and region masks. We incorporate the use of a convolutional deep belief network to learn features from greyscale, clean fingerprint images. We also show that reconstruction performed by the learnt network works as a suitable enhancement of the fingerprint, and hierarchical probabilistic inference is able to estimate overall fingerprint structures as well. Our approach performs better than Gabor-based enhancement and short time Fourier transform-assisted enhancement on images it was trained on. We further use information from the learnt features in first layer, which are short and oriented ridge structures, to extract the orientation field, frequency image, and region mask of input fingerprints. Copyright 2014 ACM.
暂无评论