Designing an effective channel attention mechanism implores one to find a lossy-compression method allowing for optimal feature representation. Despite recent progress in the area, it remains an open problem. FcaNet, the current state-of-the-art channel attention mechanism, attempted to find such an information-rich compression using Discrete Cosine Transforms (DCTs). One drawback of FcaNet is that there is no natural choice of the DCT frequencies. To circumvent this issue, FcaNet experimented on ImageNet to find optimal frequencies. We hypothesize that the choice of frequency plays only a supporting role and the primary driving force for the effectiveness of their attention filters is the orthogonality of the DCT kernels. To test this hypothesis, we construct an attention mechanism using randomly initialized orthogonal filters. Integrating this mechanism into ResNet, we create OrthoNet. We compare OrthoNet to FcaNet (and other attention mechanisms) on Birds, MS-COCO, and Places356 and show superior performance. On the ImageNet dataset, our method competes with or surpasses the current state-of-the-art. Our results imply that an optimal choice of filter is elusive and generalization can be achieved with a sufficiently large number of orthogonal filters. We further investigate other general principles for implementing channel attention, such as its position in the network and channel groupings. Our code is publicly available at (https://github.com/hady1011/OrthoNets)

This paper presents a new services paradigm for virtual world crawler interaction that is co-operative and exploits information about 3D objects in the virtual world. Our approach supports analyzing redundant information crawled from virtual worlds in order to decrease the amount of data collected by crawlers, keep search engine collections up to date, and provide an efficient mechanism for collecting and searching information from multiple virtual worlds.

Automated commercial detection can be performed by matching features extracted from commercials or by detecting embedded codes that are hidden within the commercial. In both cases, it is necessary to create a database of known commercials that contain the information necessary for detection. In this paper, we present an automated technique for locating previously unknown commercials by continuously monitoring broadcast television signals. Our system has two components: repeated video sequence detection, and feature-based classification of video sequences as commercials or non-commercials. Our system utilizes customized temporal video segmentation techniques to automatically partition the digital video signal into semantically sensible shots and scenes. As each frame of the video source is processed, we extract auxiliary information to facilitate repeated sequence detection. When the video transition marking the end of the shot/scene is detected, we are able to rapidly locate all previous occurrences of the video clip. In order to classify video sequences as commercials or non-commercials, we extract a number of features from each video sequence that characterize the temporal and chromatic variations within the clip. We have evaluated three classification approaches using this information and have consistently achieved over 90% accuracy identifying new commercials and non-commercials as they are broadcast.

Broadcasters often reuse video clips while reporting news stories. New clips are added to the story as it develops replacing old clips. A subset of these video clips may be replayed as follow up stories are investigated. This pattern of video repetition can be used to impose a story level organization on video clips that are broadcast by one or more news channels. In this paper, we describe an automated system for detecting and tracking repeated video clips in news broadcasts. We begin by performing temporal video segmentation to divide the video into shots and scenes. As each frame of the video source is processed, we extract low-level video features that are used to perform repeated sequence detection in realtime. Our matching algorithms have been adapted to recognize partial clip reuse while remaining robust to minor variations in the video source. Our system then builds a set of shots relevant to the news story being tracked, called a story core, and identifies new story episodes by detecting occurrences of shots belonging to this set. The story core is allowed to evolve dynamically over time to account for additional video footage as it becomes available. This technique was evaluated on a 24-hour broadcast of CNN News channel where we were very successful locating and tracking video clips associated with selected news stories.

Multiscale image analysis has been used successfully in a number of applications to classify image features according to their relative scales. As a consequence, much has been learned about the scale-space behavior of intensity extrema, edges, intensity ridges, and grey-level blobs. We investigate the multiscale behavior of gradient watershed regions. These regions are defined in terms of the gradient properties of the gradient magnitude of the original image. Boundaries of gradient watershed regions correspond to the edges of objects in an image. Multiscale analysis of intensity minima in the gradient magnitude image provides a mechanism for imposing a scale-based hierarchy on the watersheds associated with these minima. This hierarchy can be used to label watershed boundaries according to their scale. This provides valuable insight into the multiscale properties of edges in an image without following these curves through scale-space. In addition, the gradient watershed region hierarchy can be used for automatic or interactive image segmentation. By selecting subtrees of the region hierarchy, visually sensible objects in an image can be easily constructed.

This work provides a variational framework for fusing range and intensity data for recovering regularized surfaces. It is shown that this framework provides natural boundary conditions for the shape-from-shading problem, results in a new shape-from-shading formulation in the absence of range data, and provides a new fusion paradigm when range data is incorporated. The approach is demonstrated on simulated range and intensity images; error analysis with respect to the ground truth surface is presented. It is shown that the formulation performs well even in very noisy images.

MRIs are effective for non-invasively imaging the interior of the human brain. Due to the large amount of data associated with typical MRI sessions, manual segmentation of the images of the human brain is prohibitive except in isolated cases. The various imaging and contrast artifacts common to MRIs, however, make automatic segmentation difficult. A segmentation algorithm incorporating multi-channel MRI data is described; this approach utilizes the variational calculus formulation to simultaneously compute piecewise smooth estimates of each channel, as well as a continuous "edge process" common to all the channels.>

Two methods for identifying and analyzing the multiresolution behavior of ridges and valleys in grey-scale images are described. The first method uses the tools of differential geometry to focus on local image behavior. The resulting vertex curves mark the tops of ridges and bottoms of valleys in an image. The second method focuses on the global drainage patterns of rainfall on a terrain map. The resulting watershed boundaries also identify the tops of ridges and bottoms of valleys in an image. By following these two geometric representations through scale space, the authors build resolution hierarchies on ridges and valleys in the image that can be utilized for interactive image segmentation.>

The authors present the intensity axis of symmetry (IAS) method for describing the shape of structures in grey-scale images. They describe the spatial and intensity variations of the image simultaneously rather than by the usual two-step process of using intensity properties of the image to segment an image into regions and describing the spatial shape of these regions. The result is an image shape description that is useful for a number of computer vision applications. The method relies on minimizing an active surface functional that provides coherence in both the spatial and intensity dimensions while deforming into an axis of symmetry. Shape-based image segmentation is possible by identifying image regions associated with individual components of the IAS. The resulting image regions have geometric coherence and correspond well to visually meaningful objects in medical images.>

The failure of well-developed techniques for grey-level scalar images when directly applied to remove impulse noise with median filters in color vector images is illustrated. Pertinent criteria to remove noise in color images with median filters and reconstruct the perceptually optimum output of the median filters are derived and justified. The experiments demonstrate noise suppression and minimization of color distortion.>

A fundamental approach for providing an image description in terms of visually sensible image regions is described. It involves a) the representation of the image by a that captures image information and then b) the definition of a hierarchy of components of that by the order of annihilation of those components as the image is continuously simplified by lowering the scale. The information-capturing essential structure is chosen so that image regions are associated with each component during the image simplification. To guarantee image simplification, successive Gaussian blurring is chosen as the means of scale lowering. We argue that an that describes shape in both the spatial and intensity dimensions will produce an image description most likely to be useful for computer or human specification of image objects. In particular, we suggest that the intensity axis of symmetry (IAS) satisfies all desirable criteria for an structure. With such shape-based structures the approach of image description via annihilation under image simplification becomes a very attractive paradigm.