Vision, Graphics, and Visualization @Microsoft Research

Building Representation in Oblique-view Maps of Modern Urban Areas

This work presents a technique for creating oblique view maps of urban areas. We identify and apply cartographic and cognitive principles to develop a solution in the context of state of art Geographic Information Systems. The gap in the ability of these systems to render three-dimensional buildings into maps is addressed. At the core of our solution is a building façade modeling approach that supports varying degrees of abstraction. This is achieved by introducing a concept of “façade waveforms" and representing building façades as combinations of thesewaveforms. A Fourier series approximation of the waveforms is used during the rendering processes resulting in an elegant solution to anti-aliasing. The formulation retains the semantic information in the representation that enables meaningful extensions like night time façade generation. The solution is implemented as a pixel shader and therefore leads to a large reduction in texture memory requirement compared to existing building rendering techniques. Additionally, in the case of web based systems there is significant reduction in bandwidth requirement.

A Practical Approach to Image-Guided Building Façade Abstraction

This paper presents a technique for generating abstract representations of buildings in modern

urban areas. The solution is developed in the context of cartography, however it has applications in architectural renderings and artistically rendered games where simple stylized depictions of buildings are required to create an urban ambience. The abstraction of buildings is achieved by determining the dominant colors and primary periodic features of a building from photographs/ textures of building facades. A parametric model of building facades as waveforms, based on Fourier series, is used to approximate the façade structure. The values of the parameters (coefficients) of the waveforms are derived from the images of building facades. The periodic nature of the facades is exploited to optimize the representation. The technique works on both hand designed facade textures and photographs of buildings.

Purpose-Driven Navigation

Navigation in the context of digital maps is associated with a sequence of pan and zoom operations that lead to a specific destination. In this work, we propose creating rich navigational schemes by augmenting the existing concept of navigation with knowledge of purpose behind it. The proposed technique enables support of navigational interactions like ``scan region'' and ``explore neighborhood''.
> People involved: Neeharika Adabala, Kentaro Toyama

Woodcut Maps

We are developing methods to take a vector representation of map features as input and to render a map in the style of hand-created maps, such as antique woodcut maps. The patterns are created with procedural approaches that depend on the semantic labeling on each vector element, with, for example, mountain ranges drawn differently from coastlines. 
> People involved: Neeharika Adabala, Kentaro Toyama 
> Links: A poster paper was presented at SIGGRAPH 2005:http://www.siggraph.org/s2005/main.php?f=conference&p=posters 

The Digital Heritage Narratives Project

We are exploring ways to preserve and showcase the richness and detail of both tangible and intangible aspects of heritage. For example in the "Sri Andal Temple Narrative" project, we weave together multiple technologies into a single compelling narrative about the legend of Godess Andal and the Srivilliputhur temple complex.

Multiple Kernel Learning

Support Vector Machines (SVMs) are tools in machine learning that are used for fundamental tasks such as classification and regression. They find applicability indiverse areas ranging from vision to software engineering to natural-language processing. The success of SVMs in these areas is dependent on the choice of a good kernel, one that is often specified and fixed in advance. However, constructing and hand-tuning kernels can be difficult as can selecting and combining appropriate features.

Multiple Kernel Learning (MKL) addresses this problem by learning an SVM kernel from training data. The kernel is therefore learnt and optimized for the specific problem being considered. We are interested in various aspects of MKL including formulation, optimization and applications.

Object Recognition

The goal of object recognition is to recognize and detect objects of interest in images of everyday scenes. This is one of the most challenging problems in computer vision for a variety of factors including variations due to camera viewpoint and illumination changes, large intra category variation and small inter category variation as well as variations due to non-rigid object deformations. We are interested in coming up with features that can strike the right balance between discriminative power and invariance. We are also interested in the design of learning algorithms that can be trained reliably given the small amounts of labeled data available and which are also efficient at run time.

Texture Analysis

Humans use texture as a fundamental cue while visually analyzing the world. Yet, a rigorous mathematical and algorithmic definition remains elusive. We are interested in the various ways in which texture information can be exploited in computer vision tasks including recognition, synthesis and segmentation. In particular, our focus has been on the classification of materials in non-calibrated images. We have designed various features based on filter banks, patches and fractals and are interested in how these can be used to statistically describe textures.

Reading Text in Images of Natural Scenes

Our surroundings contain a wealth of interesting textual information. Having such text automatically read and translated for us can be very useful in many situations. The proliferation of mobile cameras has enabled us to capture this text easily in the form of images. However, reading text in images of natural scenes is a hard problem. Optical Character Recognition (OCR) solutions can not be applied out of the box as OCR systems are unavailable for many popular languages and one also has to deal with non-standard font and background, viewpoint variations and low resolution images. We are interested in various aspects of the problem including efficient detection, segmentation and recognition.

Predicting Facial Attractiveness

We are interested in predicting a person’s facial attractivenessin a given image. Generalized notions of beauty are subjective. However, an individual’s or group’s notion of beauty is often consistent and can be learnt. In the special case of faces, recent research suggests that there might even be a common, universal perception of beauty. Various factors, ranging from the evolutionary to the social and cognitive, have been attributed to explain the consistency in ratings between human subjects. Given training data in the form of photographs of faces along with their attractiveness ratings, our goal is to come up with features and a regression function which can help predict facial attractiveness in new images.

转载自http://research.microsoft.com/en-us/groups/vgv/