· · · Conference Proceedings hosted by TU Delft Library

Sustainable design is an approach that seeks to adopt an ethic of the future, where the vision of the solutions is based on a temporal and spatial perspective that is predominantly long-term and global. Design is characterized by its projective and ambivalent nature, and therefore a conscious effort to anticipate the outcomes of design intentions is crucial. Consequently, all design is inherently laden with uncertainty, doubt, and specifically in some technology-driven design projects - contradictions and controversies. Typically, such uncertainties and contradictions are not considered during the initial phase, since the main goal at this phase is to simplify the problem, and therefore these anomalies are often omitted, as they are seen to be outside the boundaries of the design problem. How can designers consider the uncertainties and contradictions during conceptualization, as well as consider the benefits resulting from their design proposals? Designers in their sustainable design practice must consider (1) the multiple objectives and criteria; (2) the multiple users and user preferences; (3) the multiple design alternatives; (4) the complex changing global situation; and (5) the knowledge from the various disciplines comprising the design project. A collective systems thinking approach to design addresses these concerns. Consequently, the theoretical basis of the precautionary principle is directly in line with this approach to design. This presentation will discuss the epistemological and practical implications of the precautionary principle for design in this context.

Until now, most of the computer games are played with a controller. But there is a big difference between pushing a button to jump and jumping yourself. Currently we are developing new tools which allow players to get rid off controllers and play games using intuitive body movements and poses. In this demo, we illustrate real-time pose tracking and recognition applied to pose-driven spatial game.

We propose a demonstration of our method for video shot boundary detection that is presented in our paper that involves clear presentation of the method and idea, together with the illustration of the results in real time. The proposed method in our paper utilizes local image features extracted around keypoints and analyzes their spatial structure. Thus, we demonstrate how those keypoints’ spatial distribution varies at each frame as the video progresses.

This work demonstrates a new approach to video temporal segmentation into scenes. The utilized technique is based on an audio-visual extension of the well-known method of the Scene Transition Graph (STG). This multi-modal extension exploits both low- and high-level audio-visual descriptors to construct distinct STGs. These STGs are employed into a probabilistic framework that is used for estimating a confidence value on each shot boundary also being a scene boundary. Finally, the thresholding of these confidence values generates the set of experimentally estimated scene boundaries. In this demo both the scene segmentation outcome and some intermediate features that lead to it are demonstrated.

The demonstration presents the operator console of the framework for detecting events in a parking lot. The console application connects safely through the Internet to the running framework installed in Gdansk, Poland, receives video streams and event data from it and allows data visualization. The installation consists of multiple cameras, including fixed and PTZ ones. PTZ cameras are able to follow moving objects automatically. The place and time of vehicles stopping in a parking lot is the main event detected by the framework. Other events include detection of vehicles stopping outside of a parking place, persons entering/leaving a building or walking along the road and detection of vehicle leaving and entering the building.

This paper discusses an approach, which allows classifying videos by frequencies. Many videos contain repetitive activities like walking, swimming, or playing table tennis. These activities usually repeat with a certain frequency, which can be seen as a feature of cyclic motion. So determining this feature can help to classify videos with repeating movements properly. In this paper we explain how to find out the right frequencies for video clips and how to use them for classifying. The main method handles series of image moments as a function in order to transform this function into the frequency domain via FFT. Techniques proposed in this paper are tested with own and with external video data. Thus it can be pointed out how the system handles different data types and data qualities.

Removing noise effectively without destroying important fine structure is a challenging problem in image processing. Among many techniques, diffusion method is shown to be attractive due to its rigorous mathematical background and efficiency. In this paper we propose a well-balanced diffusion method by redefining the diffusion and regularization function. Results obtained from processing synthetic and real images demon- strate that our new method can obtain better performance in terms of removing noises without destroying detailed features of images.

The paper presents a spatio-color classification in a chrominanceluminance space related to the dichromatic model. The unsupervised adaptive clustering is performed after the color connectedness degrees (CCD) of a color interval which embeds jointly 1) colorimetric information; 2) the probability that a given color is connected (in the image) to a set of similar colors. The chrominance CCDs are analyzed first while the luminance CCDs are studied only when necessary. Eventually, the method depends mainly on one parameter : the quantization step l. The method is evaluated quantitatively in terms of quality and compactness (number of finals colors) on the Kodak image database. As an example, this generic technique is applied to skin detection.

For face representation, the illumination effect is an important factor, in order to reduce illumination factor, we propose a facial feature location method based on improved active shape model (IASM) to reduce the drawback of intensity contrast and to enhance edge information. First, the face images are preprocessed by the proposed illumination normalization method using Gabor wavelets. Then, the location of facial features can fit more efficient and fast by the Gabor kernels reduction and the proposed feature-based weighted warping method. Based on our method, it not only obtains the better face alignment but also overcomes the active shape model (ASM) method which caused the failure for aligned target. Experimental results show that this approach can achieve in JAFFE face database with seven type facial expressions and Yale face database_B with varied illumination conditions.

Efficient media search applications can highly improve productivity in various domains. This is also the case in a broadcast environment, where large amounts of media are generated and archived. In this paper, we show how connecting a keyword thesaurus, used to annotate archived media items at the Flemish public service broadcaster in Belgium, with the Linked Open Data (LOD) cloud can greatly improve search applications. First, an algorithm is described that is used to automatically link concepts defined in a thesaurus with DBpedia, an important linking hub in the LOD cloud. The evaluation of this algorithm gives an overall accuracy of 81.64%. This is followed by an overview of features that are useful in a search application that were enabled through establishing a connection with the LOD cloud.

The spatial arrangement of visual elements of an image, i.e. the visual composition, is a research subject in the domain of visual arts which include painting, film, etc. Film experts face the problem of retrieval of visual compositions in film on a daily basis. Although, visual composition is a crucial element to consider in content-based video retrieval, little scientific effort has been invested into this problem so far. Actually, it is unclear if content-based retrieval of visual compositions is feasible. We present a user study conducted to investigate the feasibility of content-based retrieval of visual compositions as they are understood by film experts. For that reason, we create a data set derived from real world material and let the film experts evaluate the retrieval performance. The user study investigates the applicability of state-of-the-art visual features and shows differences in evaluations by film experts (test group) and computer scientists (reference group).

Automatic selection of ground-truth images is very important for the training of image classifiers, like the ones used in concept-based image retrieval. For this purpose, we propose a method which collects a sufficient number of ground-truth images based on their user-assigned tags. The semantic similarity between the tags of the images and the concept is used as a relevancy metric to classify the images in ranked lists. The system is comprised of parts of data pre-processing, WordNet-based synonym retrieval, Natural Language Processing and corpus-based semantic similarity calculations. Experimental results indicate that the proposed method is effective in collecting groundtruth data and that the training of concept classifiers based on this groundtruth leads to effective image retrieval.

In this paper, we exploit the advantages of tensor representations and propose a Supervised Multilinear Learning Model for regression. The model is based on the Canonical (CAN-DECOMP)/Parallel Factors (PARAFAC) decomposition of tensors of multiple modes and allows the simultaneous projection of an input tensor to more than one discriminative directions along each mode. These projection weights are obtained by optimizing a ϵ-insensitive loss functions which leads to generalized Support Tensor Regression (STR). The methods are validated on the problems of head pose estimation using real data from publicly available databases.

In the ongoing biodiversity crisis many species, particularly primates like chimpanzees or gorillas, are threatened. Therefore, autonomous monitoring techniques become more and more important to protect the remaining populations. However, the manual annotation of images and video sequences is not feasible for such a huge amount of data. Consequently, there is a high demand for automated analytical routine procedures. Recently, computer vision techniques for animal detection and identification are applied to overcome this issue. In this paper we compare several state-of-the-art algorithms for human face recognition for the very new field of primate photo identification. Besides common techniques like Eigenfaces, Fisherfaces, Laplacianfaces as well as more sophisticated approaches like Tensor Subspace Analysis and Volterrafaces, we also use a new concept for face recognition using a randomly generated projection matrix in conjunction with a classifier based on sparse representation. Our experimental results show that the Sparse Representation Classifier using a randomly generated projection matrix outperforms all the other algorithms.

This paper presents a GPU-assisted version of the LIBSVM library for Support Vector Machines. SVMs are particularly popular for classification procedures among the research community, but for large training data the processing time becomes unrealistic. The modification that is proposed is porting the computation of the kernel matrix elements to the GPU, to significantly decrease the processing time for SVM training without altering the classification results compared to the original LIBSVM. The experimental evaluation of the proposed approach highlights how the GPU-accelerated version of LIBSVM enables the more efficient handling of large problems, such as large-scale concept detection in video.

Without any doubt multimedia content has become essential in all aspects of our digital society, leading to the generation and storage of an unfathomable amount of digital media. In the Information Retrieval (IR) field the majority of research available to this date approaches all questions in terms of semantic tagging, indexing and feature extraction. Although these are all fundamental steps in the design of any IR system, we believe that also an efficient human machine interface (HMI) can significantly improve the retrieval rate success at the end-user side. In our work we developed a Zoomable User Interface integrated with semantic algorithms dealing with media content, calling it a Semantic ZUI: we believe thath this approach can help browsing multimedia files in a seamless way, providing benefits for end users.

In this paper, we propose a novel approach to perform scene segmentation of TV series. Using the output of our existing speaker diarization system, any temporal segment of the video can be described as a binary feature vector. A straightforward segmentation algorithm then allows to group similar contiguous speaker segments into scenes. An additional visual-only color-based segmentation is then used to refine the first segmentation. Experiments are performed on a subset of the Ally McBeal TV series and show promising results, obtained with a rule-free and genericmethod. For comparison purposes, test corpus annotations and description are made available to the community.

One-bit transform (1BT)-based block motion estimation (ME) has been proposed to reduce the computational complexity by using Boolean exclusive-OR matching of one-bit representations of image frames. In this paper, a novel early termination algorithm for 1BT-based ME is proposed in order to decrease the calculations of the block distortion measure. Unlike the classical early termination schemes, the proposed algorithm utilizes a new approach to reduce computations. It employs the binomial distribution based on the characteristic of binary plane which is composed of only two elements: 0 and 1. Experimental results show that the proposed algorithm keeps its peak signal-to-noise ratio (PSNR) performance very close to the full search algorithm (FSA) while the computational complexity of ME is reduced considerably.

In this paper, we present a part of surveillance centric indexing framework aimed at studying the performance of multi-feature fusion technique for indexing objects from surveillance videos. The multi-feature fusion algorithm determines an optimal metric for fusing low-level descriptors extracted from different feature space. These low-level descriptors exhibit a non-linear behaviour and typically consist of different similarity metrics. The framework also includes a motion analysis component for the extraction of objects as blobs from individual frames. The proposed framework, in particular the multi-feature fusion algorithm is evaluated against kernel machines for indexing objects such as car and person on AVSS 2007 surveillance dataset.

Motion compensated background subtraction is a common method for generic video object segmentation. The basic principle behind this approach is that the objects move differently to the global motion of the background. Hence a motion compensation of disclosed background pixels in previous or subsequent frames allows for the mosaicing of the background. This paper presents a stochastic method for the background mosaicing in dynamic scenes containing unconstrained camera motion, zoom, rotation and even (weak) lens distortion. Thereby we countervail inaccurate motion compensation and exploit the correlation of adjacent background mosaics. That way, the generation of the background mosaics can be performed very efficiently, making it suitable for real-time applications. A quantitative evaluation illustrates the validity of this approach.