Demo Program

Demo Chair: Matthew Ma (Email: mattma@ieee.org)
Demo Logistics: Rebecca Chau (Email: ceb@ctshk.com)

The ICPR 2006 will hold demo sessions in conjunction with conference meetings to encourage discussions and information exchanges among attendees.

Demo Schedule

List of Registered Demos

• 3D Geometry Reconstruction and Statistical Shape Modeling
  Zuse Institute Berlin, Germany
• Fast Linear Feature Detection using Multiple Directional Non-Maximum Suppression
  CSIRO Mathematical & Information Sciences NSW 1670, Australia
• Robust Head Tracking and Illumination Robust Facial Expression Recognition System
  Department of Computer Engineering, POSTECH, KOREA
• Gesture Recognition using Temporal Templates
  Department of Computer Engineering, POSTECH, KOREA
• Real-Time Document Image Retrieval Using Web Cameras
  Graduate School of Engineering, Osaka Prefecture University, Japan
• Stereo Vision, Spherical Video, and High Frame-rate cameras
  Point Grey Research, Inc., Vancouver, BC, Canada

Abstracts of Registered Demos

3D Geometry Reconstruction and Statistical Shape Modeling

Abstract:

Generating geometric models of anatomical objects from 2D or 3D image data is a prerequisite for many tasks in computer-aided medical diagnostics and planning. Each imaging modality (typically CT, MRT, US, X-ray, etc.) exhibits its own specific properties to be considered in the feature extraction or segmentation process. Manual methods are time-consuming, prone to errors and hardly reproducible. Incorporating a-priori knowledge about the object and the data to be segmented seems to be a feasible way to automate this task.

In this demo, we will present a software, based on the 3D visualization and volume modeling system Amira, which allows to generate and utilize statistical models of arbitrary 3D shapes. Amira already offers a variety of image filtering methods, interactive segmentation tools and a surface reconstruction algorithm for creating surfaces as the basis for the training of a statistical model. The focus of this presentation will be on how to establish correspondence across different shapes, register different shapes to a common reference coordinate system and perform statistical analysis.

Statistical shape models can be applied to a range of different applications: segmentation of 3D image data (especially low-contrast image data), 3D reconstruction from 2D data (X-ray images), reconstruction of pathological, incomplete or sparse surface data. Different anatomical models and their application will be presented in this demo.

^ Back to top

Fast Linear Feature Detection using Multiple Directional Non-Maximum Suppression

Abstract:

The capacity to detect linear features is central to image analysis, computer vision, and pattern recognition; and it has practical applications in areas such as neurite outgrowth detection, retinal vessel extraction, skin hair removal for malonoma detection, plant root analysis, and road detection. Linear features detection often represents the starting point for segmentation and image interpretation. Here, we present a new algorithm for linear feature detection using multiple directional non-maximum suppression. Given its low computational complexity, the algorithm is very fast. We show on several examples that it performs remarkably well in terms of sensitivity and continuity of detected features.

^ Back to top

Robust Head Tracking and Illumination Robust Facial Expression Recognition System

Abstract:

Active Appearance Model (AAM) is a well-known model that can represent a non-rigid object effectively. However, the AAM often fails to fit the input image under the severely changing illumination condition, because it uses the fixed appearance basis vectors that are usually obtained in the training phase. To overcome this disadvantage, we propose an adaptive AAM that updates the appearance basis vectors with the input image using incremental principal component analysis (PCA). We also propose a layered generalized discriminant analysis (GDA) classifier which combines shape and appearance information to improve the recognition performance of person independent facial expressions. This system runs at 6~8 frames per second on a Pentium 3.2GHz PC and the average expression recognition performance is 90.1%. We also propose another demo which represents a robust head tracking using 3D cylindrical model. Eye detection and active appearance model are used for initializing the 3D cylinder model. Iterative re-weighted least squares technique has been used for fitting algorithm and dynamic template technique approach can be robustly tracking about gradual lighting changes. Our system runs in real-time about 10 fps and can robustly track the slowly changing head motion. Possible tracking ranges are -90 degrees to 90 degrees horizontally and -60 degrees to60 degrees vertically.

^ Back to top

Gesture Recognition using Temporal Templates

Abstract:

We present a gesture recognition system using a stereo camera for the intelligent robot in the typical indoor environment. The gesture recognition system represents a specific motion by MEI (Motion Energy Image) and MHI (Motion History Image) view-specific temporal templates, where MEI is a binary representation of which motion has occurred in an image sequence and MHI is a scalar-valued image whose intensity is a function of regency of motion. The gesture recognition system classifies each motion using GDA ensemble via two steps: spotting the meaningful gesture section in the image sequence and recognizing the meaningful section into one of 10 specific gestures. The gesture recognition system runs at 6~8 frames per second on a 3.2GHz PC and shows an 80% accuracy of segmenting image sequence to the meaningful gesture section and a 90% accuracy of the gesture recognition.

^ Back to top

Real-Time Document Image Retrieval Using Web Cameras

Abstract:

Camera-based document image retrieval is a task of searching document images from the database based on query images captured using digital cameras. We have already proposed a method called "locally likely arrangement hashing (LLAH)" which enables us a fast (about 100 msec.) and accurate (more than 94%) retrieval from a large database (including 10,000 document images). In this presentaion, we propose a real-time document image retrieval based on LLAH. The proposed method repeats retrieval and display the result using a web camera. The system currently achieves 7 fps for the retrieval from the database of 20,000 images. As an application of the real-time retrieval, we also propose a method of augmented reality for documents. This method is to superimpose relevant information onto the camera-captured image naturally by using the parameters of perspective transformation calculated as a subsidiary result of retrieval. We consider that the results indicate a new possibility of document images as media of displaying information.

^ Back to top

Stereo Vision, Spherical Video, and High Frame-rate cameras

Abstract:

Point Grey Reseach will be demonstrating a variety of advanced cameras suitable for computer vision reseach. These include stereo vision, spherical/panoramic video and high frame-rate cameras.

Bumblebee2 Stereo Vision camera - this is a real-time stereo camera that runs at speeds up to 48 frames/second. The camera uses high quality Sony CCD sensors and is available in color or B&W models. It transmits images to the host computer over a 1394 FireWire interface. It is factory calibrated and stays in calibration. The Triclops C/C++ stereo SDK is provided with the camera. Real-time rectified and disparity images will be demonstrated, as well as live 3D point clouds.

Ladybug2 Spherical Video system - this is a video system that captures 75% of a full sphere of video data at speeds up to 30 frames/second. The camera is a cluster of six sensors tightly packaged to minimize the effects of parallax. This number of sensors allows the system to deliver 4.7 megapixels of image data with every frame. The Ladybug software development kit allows geometric processing from individual sensor images as well as a suite of routines for rendering panoramic images in various formats. The Ladybug2 camera will be demonstrated running live, as well as pre-recorded outdoor images. Free DVDs with sample data will be available.

Dragonfly Express high-framerate VGA camera - this camera produces images at 200 frames/second with VGA resolution over off-the-shelf 1394B interface technology. This camera is well-suited for high-speed tracking and provides impressive performance for a modest price. 200 frames/second operation will be demonstrated.

Point Grey Research makes a range of high quality 1394 cameras suitable for vision research. Resolutions range from 640x480 to 1600x1200. PGR cameras auto-synchronize with other cameras on the 1394 bus, making them ideal for multi-camera applications. The cameras are feature-packed and completely configurable via computer over a single interface. They are Linux compatible. A variety of cameras will be on-hand for demonstration to interested parties.

Please visit http://www.ptgrey.com for further information on Point Grey Research cameras.

^ Back to top