Understanding object motions and transformations is a core problem in computer science. Modeling sequences of evolving images may provide better representations and models of motion and may ultimately be used for forecasting or simulation. Diagrammatic Abstract Reasoning is an avenue in which diagrams evolve in complex patterns and one needs to infer the underlying pattern sequence and generate the next image in the sequence. For this, we develop a novel Contextual Generative Adversarial Network based on Recurrent Neural Networks (Context-RNN-GANs), where both the generator and the discriminator modules are based on contextual history and the adversarial discriminator guides the generator to produce realistic images for the particular time step in the image sequence. We employ the Context-RNN-GAN model (and its variants) on a novel dataset of Diagrammatic Abstract Reasoning as well as perform initial evaluations on a next-frame prediction task of videos. Empirically, we show that our Context-RNN-GAN model performs competitively with 10th-grade human performance but there is still scope for interesting improvements as compared to college-grade human performance.

Topic-models for video analysis have been used for unsupervised identification of normal activity in videos, thereby enabling the detection of anomalous actions. However, while intervals containing anomalies are detected, it has not been possible to localize the anomalous activities in such models. This is a challenging problem as the abnormal content is usually a small fraction of the entire video data and hence distinctions in terms of likelihood are unlikely. Here we propose a methodology to extend the topic based analysis with rich local descriptors incorporating quantized spatio-temporal gradient descriptors with image location and size information. The visual clips over this vocabulary are then represented in latent topic space using models like pLSA. Further, we introduce an algorithm to quantify the anomalous content in a video clip by projecting the learned topic space information. Using the algorithm, we detect whether the video clip is abnormal and if positive, localize the anomaly in spatio-temporal domain. We also contribute one real world surveillance video dataset for comprehensive evaluation of the proposed algorithm. Experiments are presented on the proposed and two other standard surveillance datasets.

Cricket broadcast video analysis has had difficulty identifying aspects of the content such as the type of batting stroke or the direction of the field played towards. Here we construct a composite feature combining Optical flow analysis along with camera view analysis to model the type of shots played. The work first presents an improved camera shot analysis based on learning parameters from a small supervision set. This splits the broadcast video into shots which are combined into balls and, the segment where the batsman is playing the stroke is identified. After that optical flow analysis is used to determine the direction of the stroke with an accuracy of 80 percent.

Animation plays a key role in adding realism to any graphical environment. In this work we present a technique for animating multiple avatars in a virtual environment designed to support remote collaboration between distributed work teams in which users are represented by avatars. There will be long periods of time when the user is not actively controlling the avatar and working on his official task. We need to animate his avatar with a `working-at-desk' animation that should be non-looping and sufficiently random for a single avatar as well as between multiple avatars to appear realistic. We present a technique for generating multiple head motions using (a) nonlinear dimensionality reduction and (b) gaze space images to control the motion. Our technique can automatically generate long sequences of motion without any user intervention. We present results from synthetic data.

Metaphors being at the heart of our language and thought process, computationally modelling them is imperative for reproducing human cognitive abilities. In this work, we propose a plausible grounded cognitive model for artificial metaphor acquisition. We put forward a rule-based metaphor acquisition system, which doesn't make use of any prior 'seed metaphor set'. Through correlation between a video and co-occurring commentaries, we show that these rules can be automatically acquired by an early learner capable of manipulating multi-modal sensory input. From these grounded linguistic concepts, we derive classes based on lexico-syntactical language properties. Based on the selectional preferences of these linguistic elements, metaphorical mappings between source and target domains are acquired.

Occlusion is often thought of as a challenge for visual algorithms, specially tracking. Existing literature, however, has identified a number of occlusion categories in the context of tracking in ad hoc manner. We propose a systematic approach to formulate a set of occlusion cases by considering the spatial relations among object support(s) (projections on the image plane) with the detected foreground blob(s), to show that only 7 occlusion states are possible. We designate the resulting qualitative formalism as Oc-7, and show how these occlusion states can be detected and used effectively for the task of multi-object tracking under occlusion of various types. The object support is decomposed into overlapping patches which are tracked independently on the occurrence of occlusions. As a demonstration of the application of these occlusion states, we propose a reasoning scheme for selective tracker execution and object feature updates to track multiple objects in complex environments.

ldquoGhostsrdquo arise in traditional background subtraction when an object starts to move, causing the exposed background to be labelled as a ghost foreground. With background model updates, the ghost may disappear after some time, but removing ghosts immediately is crucial for identifying starts, and also for improving downstream tasks like tracking, object recognition and activity analysis. Here we propose a lagged background subtraction process, where the background model is computed in reverse after a lag of k frames, resulting in a ldquoreversed-timerdquo background model. We present an algorithm that handles disparities between the forward and backward foreground blobs, and show how ghosts can be reliably eliminated, potentially permitting single-frame latency in reliably detecting starts and identifying stops. We present theoretical results that the bidirectional model results in lower false positives (such as ghosts) compared to either directional model, and develop an algorithm for identifying ghosts.

Traditionally it has been assumed "Design Specifications" are given a priori, and constitute the primary constraints on the process of search that is design. This view is under challenge in emergentist accounts of design, where it is seen that for well-understood functional needs, experienced designers are able to come up with good designs very quickly. It is hypothesized that this is possible because search is minimized using novel functional constraints that emerge from experience. These emergent aspects are difficult to model in a computational framework, and this work is a preliminary attempt in this direction. "Well-understood functions" are assumed to be quantifiable in terms of some performance metrics, which permits us to identify regions of high functional validity as emergent constraint regions in the design space. In addition, designers often change the design space itself, and negotiate the initial specs in many ways. We show that small changes in the design space may result in large changes in this mapping, which is why such emergent knowledge is limited to a specific embodiment. By introducing such measures into future solid modeling systems, it may reduce the human designer's search to the more ill-posed aspects of the problem.

In this paper we describe a method for the synthesis of robotic controllers using evolutionary techniques. A modified version of the recurrent neural network used for controlling the robots is evolved using genetic algorithm using the variable length genotype approach where the genotype encodes the network. It has been discovered that the separation of modalities in the network like vision and touch, for the first few layers helps in faster evolution as well as in the development of faster controller networks. The structure of the network that emerges from this kind of evolution is similar to brain-like networks. Performance of plastic versus non-plastic individuals has also been explored. Gene-blocking technique has been used for developing several behaviors in the same evolution cycle. The final controller developed at the end of the evolutionary process was tested on a Khepera.

Reactive robot behavior has been claimed as a model of memoryless intelligence. Using results highlighting the cyclic conflicts in such systems, we throw light on similar phenomena in biology such as human amnesic behavior. We present a formal characterization of cycles in the temporal sequence of memoryless behaviors and show that attempts to eliminate cycles result in less powerful behavior systems. Conversely, more powerful behavior systems are more likely to exhibit cyclicity. Such systems can be viable in performing simple tasks only by using the world as an external memory. Our results also provide useful insights into a better design of strategies for the treatment of amnesics.

This research focuses on robot behaviours which use minimal communication and rely mostly on changes in the environment as their cue for action. The behaviour-based paradigm for building autonomous robots has recently become very popular because of its successes, use of the world as an external memory and replacement of classical planning by agent-environment dynamics. However there are no metrics for evaluating and improving the behaviour spaces. Our aim here is to bridge this gap. We define novel metrics (power, usefulness, flexibility, modularity and scalability) and investigate the properties of behaviour spaces using them. We use these metrics to present results on modifications to individual behaviours and addition of new behaviours to the behaviour spaces. We discuss the case of a behaviour-based robot operating in kitchen to illustrate the significance of our metrics and discuss how the utility of our metrics remains valid in other behaviour representations.

Owing to the exponential costs of path planning in a continuous or graded cost environment, robot motion planning traditionally makes the "perfect obstacle assumption" and divides workspace into perfect obstacles and perfect freespace, although in practice such a black-and-white distinction is rare. Under the above definition, however, many finite cost regions can also be shown to be perfect, substantially reducing the computational costs. We present a linear-time algorithm for deciding whether an obstacle is perfect in the convex case, and a genetic algorithms approach in the nonconvex case. When the obstacle is not perfect, we identify a measure of the degree to which it approximates a perfect obstacle. Identifying perfect obstacles helps avoid situations where it may be possible to push aside an obstacle, or climb a hillock, for example.

Feature-driven CAPP is the dominant model for computer-aided process planning today. Despite more than a decade of effort, there are no general methods for defining or identifying features, and consequently CAPP approaches in full generality could not been developed. An alternative to feature-driven CAPP is direct tool access-based models for process planning, which rely on the fundamental geometric information in the CAD model to develop the process plan. In this paper we check the accessibility of part surfaces from different tool access directions, and use this to generate the process plan for rotational parts (turning) and for milling. Features are still useful, especially for communicating with the shop floor user, the part faces are clustered into features for the purpose of process detailing. Thus features may be seen as being driven by the process plan and not the other way round.

Qualitative spatial reasoning has been used for representing landmarks, path and grasp planning, shape recognition etc, but has generally not been extended beyond 2-D. In this paper, we develop a 3-D spatial reasoning algebra, using qualitative versions of (a) Gibbs-Heaviside vector calculus with homogeneous coordinates, and (b) screw calculus motions that map one object frame onto another. By comparing the results of composition in one, two and three dimensions, we find that the degree of uncertainty is much higher in 3-D. We demonstrate the utility of a qualitative formulation in a simple robot spatial task with imprecise instructions.

Traditional robot motion planning makes the "perfect obstacle assumption" and divides workspace into perfect obstacles and perfect freespace. A region of the workspace is a perfect obstacle if the cost of traversing it is such that there always exists a more efficient path external to the region. Identifying perfect obstacles helps avoid situations where it may be possible to push aside an obstacle, or climb a hillock, for example. We present a linear-time algorithm for determining perfect obstacles in the convex case, and a genetic algorithms approach in the nonconvex case. When the obstacle is not perfect, we identify a measure of the degree to which it approximates a perfect obstacle. This can be a pre-condition to applying standard motion planning techniques.

A novel approach to building qualitative models for vision based on the functionally relevant attributes of each object is presented. The spatial discretization used is derived systematically, and is complete under the conditions of tangency, no-contact, and overlap. Each object is decomposed into simpler qualitative generalized cylinders, and the join operations are specified by a new three-dimensional qualitative model for space. Results are shown for applications in two and three dimensions.>

The authors present a method to determine whether or not a part is machinable on a three-axis milling machine by consideration of tool motion. The method focuses on determining sets of candidate machining directions, called the basis set, and testing if the process plan generated from any particular set of machining directions succeeds in machining all of the part features. This constitutes a machinability test that can be used at design. The range of parts that are successfully handled include all simple polyhedrons and simple three-dimensional curved faces such as cylinders, cones, and spheres.>

The development of fast recursive methods for computing manipulator inverse dynamics has made possible open loop control strategies. However, for these strategies to work, an accurate plant model is required. Two key components of the plant are frictional terms and link/load inertias. This paper shows how these components may be computed using force and moment sensing.