DCVISIGRAPP 2016 Abstracts

Full Papers

Paper Nr:	1
Title:	Scene Analysis from Color and Depth Data based on Surface Fitting
Authors:	Giampaolo Pagnutti
Abstract:	Scene segmentation is a difficult problem, that is often hard to tackle with satisfactory results if color information alone is taken into account. My research proposes novel iterative segmentation methods, that jointly use color and depth data combined with a 3D surface approximation scheme. First a set of multi-dimensional vectors is built from color and geometry information. Normalized cuts spectral clustering can then be applied to subdivide the scene into a predetermined number of segments. I propose an iterative region splitting scheme where binary segmentation is used at each step, and segments that do not represent a single surface are split again into smaller regions, until an optimal segmentation is obtained. I also propose a region merging approach where I initially over-segment the scene into a large number of small segments, then I apply an iterative merging procedure that recombines the adjacent compatible segments into the regions corresponding to the actual objects. In both methods, a Non-Uniform Rational B-Spline (NURBS) model is fitted on each of the computed segments. The accuracy of the fitting is considered as a measure of the plausibility that the segment represents a single surface or object, and this is the main criterion adopted to continue or stop the splitting and merging procedures. Experimental results show how the proposed methods provide accurate and reliable scene segmentations, comparable to those of state-of-the-art methods on a standard dataset. Further enhancements are expected from combination of the two approaches, that is, by performing a recursive splitting procedure followed by a region merging scheme. In addition, the proposed methods show a good potential for semantic segmentation and depth map coding applications.

Paper Nr:	2
Title:	Peripheral Blood Image Analysis
Authors:	Andrea Loddo, Cecilia Di Ruberto and Lorenzo Putzu
Abstract:	The visual analysis and the counting of white blood cells in microscopic peripheral blood smears is a very important procedure in the medical field. It can provide useful information concerning the health of the patients, e.g., the diagnosis of Acute Lymphoblastic Leukemia or other important diseases. Blood experts in clinical centres traditionally use these methods in order to perform a manual analysis. The main issues of the traditional human analysis are certainly related to the difficulties encountered during this type of procedure: generally, the process is not rapid and it is strongly influenced by the operator's capabilities and tiredness; moreover, we remind that white blood cell segmentation is the most important and difficult procedure to implement in order to realize a reliable automated system. The main purpose of this work is to realize a reliable Automated System based on a multi class SVM in order to manage all the regions of immediate interests inside a blood smear: white blood cells nucleus and cytoplasm, erythrocytes and background. The innovative aspect of our work is certainly the use of a classification method to build an image segmentation system.

Paper Nr:	3
Title:	Improving Function Detecting Object for AGV
Authors:	Jittima Varagul and Toshio Ito
Abstract:	The objective of this research is to developed and improved function detecting object accurately and efficiently for Automated Guide Vehicle (AGV). Currently, the AGV is a transport vehicle widely used in manufacturing industry. This system can also reduce labor costs in industry, as well as increasing the safety of its employees from working in dangerous environments. It can move materials and equipment to various locations that work automatically according to conditions, and the directions to destination are defined, but it cannot avoid obstacles in front by itself. The computer vision with deep learning neural networks can help the AGV can see and recognize like a human. When considering the actual AGV driving situation on the path, it is desirable to be able to recognize the preceding obstacles. In order to prevent a collision the AGVs and the obstacles, which do not know the exact shape, size and color. Such real-time obstacle detection was crucial, we need to classify the obstacles that are real obstacles or fake obstacles such as a painting on the floor, then decide whether to hold or continue to be moved. As a result, this research will focus on a new technology concerning the AGV to detect obstacles accurately and efficiently by computer vision with deep learning.

Paper Nr:	4
Title:	Data Scalable Approach for Identifying Correlation in Large and Muti-dimensional Data
Authors:	Hoa Nguyen and Paul Rosen
Abstract:	Correlation is a powerful relationship measure used in science, engineering, and business to estimate trends and make forecasts. When the data is complex, large and high dimensional, providing the method that improves correlation identification is challenging. There are many visualization methods are proposed to solve these problems but they still have limitations in the accuracy and speed. Therefore, we propose four visualization techniques to solve these problems in large and multi-dimensional data. Depending on purpose of visualization tasks, best fit technique can be provided to optimize the correlation identification performance. Existing visualization methods, such as Scatterplots (SCPs) and Parallel Coordinates Plots (PCPs), are designed to be general supporting many visualization tasks, including identifying correlation. However, due in large part to their generality, they do not provide the most efficient interface, in terms of speed and accuracy, for many tasks. To improve correlation identification tasks in low level, we propose a new correlation task-specific visualization method called Correlation Coordinate Plots (CCPs). CCPs transform data into a powerful coordinate system for estimating the direction and strength of correlation between attributes. However, correlation identification task is especially challenging as well when the number of dimensions is high, leading to many potential relationships, and/or multiway dependencies are of interest. Several visualization methods have been proposed to aid the exploration of such information through the direct visualization of summary statistics. However these methods are typically limited to the study of all possible pairwise and 3-way relationships and are rather rigid to interactive exploration to low-dimensional subspace. Therefore, we propose three different visualization designs to optimize correlation identification task in large and multi-dimensional data. The first is the Snowflake Visualization, a focus+context layout for exploring all pairwise correlations. We also enhances the basic CCP interface by using principal component analysis to project multiple attributes. The second proposed design is a new interactive design for representing and exploring data relationships in PCPs. The approach exploits the point/line duality property of PCPs and a local linear assumption of data to extract and to represent relationship summarizations. This approach simultaneously shows relationships in the data and the consistency of those relationships. Our approach supports various visualization tasks including cluster analysis, mixed linear and nonlinear pattern identification, hidden pattern detection, and outlier detection, all in large data. Finally, we propose a novel technique for storing and accessing these multiway dependencies through visualization. Exploration is supported by a variety of operations placed on the complex, and interactive visualization enables flexible investigations through overview and detail views of the data. We provide various use cases, compare to the prior works and user study to demonstrate how our proposed approaches helps to explore correlation in large and high dimensional data efficiently. These results confirmed that our approaches, CCP/Snowflake, DSPCP and MultiDepViz methods outperform some current visualization techniques such as SCP, PCP, SCP matrix, Corrgram, Angular Histogram (AngHist), and UntangleMap in both accuracy and timing.