Search

In this paper we describe K-Space’s participation in TRECVid 2008 in the interactive search task. For 2008 the K-Space group performed one of the largest interactive video information retrieval experiments conducted in a laboratory setting. We had three institutions participating in a multi-site multi-system experiment. In total 36 users participated, 12 each from Dublin City University (DCU, Ireland), University of Glasgow (GU, Scotland) and Centrum Wiskunde & Informatica (CWI, the Netherlands). Three user interfaces were developed, two from DCU which were also used in 2007 as well as an interface from GU. All interfaces leveraged the same search service. Using a latin squares arrangement, each user conducted 12 topics, leading in total to 6 runs per site, 18 in total. We officially submitted for evaluation 3 of these runs to NIST with an additional expert run using a 4th system. Our submitted runs performed around the median. In this paper we will present an overview of the search system utilized, the experimental setup and a preliminary analysis of our results.

International audience; We describe in this paper a method for 3D shape indexing and retrieval that we apply on three data collections of the SHREC - SHape Retrieval Contest 2008: Stability on watertight, CAD and Generic 3D models. The method is based on a set of 2D multi-views after a pose and scale normalization of the models using Continuous PCA and the enclosing sphere. In all views we extract the models silhouettes and compare them pairwise. To compute the similitude measure we consider the external contour of the silhouettes, we extract their convexities and concavities at different scale levels and we build a multiscale representation. The pairs of contours are then compared by elastic matching achieved by using dynamic programming.

In this paper we describe K-Space participation in TRECVid 2007. K-Space participated in two tasks, high-level feature extraction and interactive search. We present our approaches for each of these activities and provide a brief analysis of our results. Our high-level feature submission utilized multi-modal low-level features which included visual, audio and temporal elements. Specific concept detectors (such as Face detectors) developed by K-Space partners were also used. We experimented with different machine learning approaches including logistic regression and support vector machines (SVM). Finally we also experimented with both early and late fusion for feature combination. This year we also participated in interactive search, submitting 6 runs. We developed two interfaces which both utilized the same retrieval functionality. Our objective was to measure the effect of context, which was supported to different degrees in each interface, on user performance. The first of the two systems was a ‘shot’ based interface, where the results from a query were presented as a ranked list of shots. The second interface was ‘broadcast’ based, where results were presented as a ranked list of broadcasts. Both systems made use of the outputs of our high-level feature submission as well as low-level visual features.

This paper presents a new method for segmentation of images into large regions that reflect the real world objects present in a scene. It explores the feasibility of utilizing spatial configuration of regions and their geometric properties (the so-called Syntactic Visual Features [1]) for improving the correspondence of segmentation results produced by the well-known Recursive Shortest Spanning Tree (RSST) algorithm [2] to semantic objects present in the scene. The main contribution of this paper is a novel framework for integration of evidence from multiple sources with the region merging process based on the Dempster-Shafer (DS) theory [3] that allows integration of sources providing evidence with different accuracy and reliability. Extensive experiments indicate that the proposed solution limits formation of regions spanning more than one semantic object.

Dublin City University (DCU) and University of Tampere(UTA) participated in the ImageCLEF 2007 photographic ad-hoc retrieval task with several monolingual and bilingual runs. Our approach was language independent: text retrieval based on fuzzy s-gram query translation was combined with visual retrieval. Data fusion between text and image content was performed using unsupervised query-time weight generation approaches. Our baseline was a combination of dictionary-based query translation and visual retrieval, which achieved the best result. The best mixed modality runs using fuzzy s-gram translation achieved on average around 83% of the performance of the baseline. Performance was more similar when only top rank precision levels of P10 and P20 were considered. This suggests that fuzzy sgram query translation combined with visual retrieval is a cheap alternative for cross-lingual image retrieval where only a small number of relevant items are required. Both sets of results emphasize the merit of our query-time weight generation schemes for data fusion, with the fused runs exhibiting marked performance increases over single modalities, this is achieved without the use of any prior training data.

In this paper we present a general framework we have developed for running and evaluating automatic image and video segmentation algorithms. This framework was designed to allow effortless integration of existing and forthcoming image segmentation algorithms, and allows researchers to focus more on the development and evaluation of segmentation methods, relying on the framework for encoding/decoding and visualization. We then utilize this framework to automatically evaluate four distinct segmentation algorithms, and present and discuss the results and statistical findings of the experiment.

This thesis considers different aspects of the utilization of contour information and syntactic and semantic image segmentation for object registration, modeling and retrieval in the context of content-based indexing and retrieval in large collections of images. Target applications include retrieval in collections of closed silhouettes, holistic w ord recognition in handwritten historical manuscripts and shape registration. Also, the thesis explores the feasibility of contour-based syntactic features for improving the correspondence of the output of bottom-up segmentation to semantic objects present in the scene and discusses the feasibility of different strategies for image analysis utilizing contour information, e.g. segmentation driven by visual features versus segmentation driven by shape models or semi-automatic in selected application scenarios. There are three contributions in this thesis. The first contribution considers structure analysis based on the shape and spatial configuration of image regions (socalled syntactic visual features) and their utilization for automatic image segmentation. The second contribution is the study of novel shape features, matching algorithms and similarity measures. Various applications of the proposed solutions are presented throughout the thesis providing the basis for the third contribution which is a discussion of the feasibility of different recognition strategies utilizing contour information. In each case, the performance and generality of the proposed approach has been analyzed based on extensive rigorous experimentation using as large as possible test collections.

In this paper we report on our work in realising an approach to video shot matching which involves automatically segmenting video into abstract intertwinded shapes in such a way that there is temporal coherency. These shapes representing approximations of objects and background regions can then be matched giving fine-grained shot-shot matching. The main contributions of the paper are firstly the extension of our segmentation algorithm for still images to spatial segmentation in video, and secondly the introduction a measurement of temporal coherency of the spatial segmentation. This latter allows us to quantitatively demonstrate the effectiveness of our approach on real video data.

In this paper, an integrated approach to fast and efficient construction of statistical shape models is proposed that is a potentially useful tool in Information Retrieval(IR). The tool allows intuitive extraction of accurate contour examples from a set of images using a semi-automatic segmentation approach. The user is allowed to draw on the scene by simply dragging a mouse over the image and creating a set of labelled scribbles for the objects to be segmented. An automatic segmentation algorithm uses the scribbles to partition the scene and extract objects’ contour. A set of labelled points (landmarks) is identified automatically on the set of examples thereby allowing statistical modeling of the objects’ shape. The main contribution of this paper is the new approach to automatic landmark identification eliminating the burden of manual landmarking. The approach utilizes a robust method for pairwise correspondence proposed originally in [1, 2]. The landmarks are used to train statistical shape models known as Point Distribution Models (PDM) [11]. Qualitative results are presented for 3 classes of shape which exhibit various types of nonrigid deformation.

This paper presents a robust and efficient method for segmentation of images into large regions that reflect the real world objects present in the scene. We propose an extension to the well known Recursive Shortest Spanning Tree (RSST) algorithm based on a new color model and so-called syntactic features [1]. We introduce practical solutions, integrated within the RSST framework, to structure analysis based on the shape and spatial configuration of image regions. We demonstrate that syntactic features provide a reliable basis for region merging criteria which prevent formation of regions spanning more than one semantic object, thereby significantly improving the perceptual quality of the output segmentation. Experiments indicate that the proposed features are generic in nature and allow satisfactory segmentation of real world images from various sources without adjustment to algorithm parameters.

This paper presents an efficient method for calculating the similarity between 2D closed shape contours. The proposed algorithm is invariant to translation, scale change and rotation. It can be used for database retrieval or for detecting regions with a particular shape in video sequences. The proposed algorithm is suitable for real-time applications. In the first stage of the algorithm, an ordered sequence of contour points approximating the shapes is extracted from the input binary images. The contours are translation and scale-size normalized, and small sets of the most likely starting points for both shapes are extracted. In the second stage, the starting points from both shapes are assigned into pairs and rotation alignment is performed. The dissimilarity measure is based on the geometrical distances between corresponding contour points. A fast sub-optimal method for solving the correspondence problem between contour points from two shapes is proposed. The dissimilarity measure is calculated for each pair of starting points. The lowest dissimilarity is taken as the final dissimilarity measure between two shapes. Three different experiments are carried out using the proposed approach: letter recognition using a web camera, our own simulation of Part B of the MPEG-7 core experiment “CE-Shape1” and detection of characters in cartoon video sequences. Results indicate that the proposed dissimilarity measure is aligned with human intuition.

In this paper we present the Qimera segmentation platform and describe the different approaches to segmentation that have been implemented in the system to date. Analysis techniques have been implemented for both region-based and object-based segmentation. The region-based segmentation algorithms include: a colour segmentation algorithm based on a modified Recursive Shortest Spanning Tree (RSST) approach, an implementation of a colour image segmentation algorithm based on the K-Means-with-Connectivity-Constraint (KMCC) algorithm and an approach based on the Expectation Maximization (EM) algorithm applied in a 6D colour/texture space. A semi-automatic approach to object segmentation that uses the modified RSST approach is outlined. An automatic object segmentation approach via snake propagation within a level-set framework is also described. Illustrative segmentation results are presented in all cases. Plans for future research within the Qimera project are also discussed.

Current commercial outdoor Mobile AR applications rely mostly on GPS antennas, digital compasses and accelerometers. Due to imprecise readings, the 2D placement of points of interest (POI) on the display can be uncorrelated with reality. We present a novel method to geo-locate a mobile device by rec- ognizing what is captured by its camera. A visual recognition algo- rithm in the cloud is used to identify geo-located reference images that match the camera’s view. Upon correct identification, fusion of sensor data from the mobile device and the server is used to es- tablish the most probable location. The corrected geo-location can then be used on the mobile device for further POI representation or as a better initialization step for further continuos tracking.

In recent years significant progress has been made in the field of visual search, mostly due to the introduction of powerful local image features. At the same time, a rapid development of mobile platforms enabled deployment of image retrieval systems on mobile devices. Various mobile applications of the visual search have been proposed, one of the most interesting being geo-localization service based on image recognition and other positioning information. This thesis attempts to advance the existing visual search system developed at Telefonica I+D Barcelona so it could be used for high precision geo-localization. In order to do so, this dissertation tackles two significant challenges of image retrieval: improvement in robustness of the detection of similarities between images and increase of discriminatory power. The work advances the state-of-the-art with three main contributions. The first contribution consists of the development of an evaluation framework for visual search engine. Since the assessment of any complex system is crucial for its development and analysis, an exhaustive set of evaluation measures is selected from the relevant literature and implemented. Furthermore, several datasets along with the corresponding information about correct correspondences between the images have been gathered and unified. The second contribution considers the representation of image features describing salient regions and attempts to alleviate the quantization effects introduced during its creation. A mechanism that in the literature is commonly referred to as soft assignment is adapted to a visual search engine of Telefonica I+D with several extensions. The third and final contribution consists of a post-processing stage that increases discriminative power by verification of local correspondences' spatial layout. The performance and generality of the proposed solutions has been analyzed based on extensive evaluation using the framework proposed in this work.