Category Archives: Archive

CGV Colloquium Friday September 18th

We have two midterm master project presentations on Friday, 18 September starting 15:45. The session will be on Zoom.

Presenter: Berend Baas

Title: Latent shape editing

Abstract: In recent years, deep learning on shapes and manifolds has been used to try and perform a variety of tasks, such as classification, deformation transfer and shape matching. This is often done through architectures such as Autoencoders or Generative Adversarial Networks, that try to learn a vector representation of training shapes, which is then used for downstream tasks.

However, current trained representations are generally poorly structured: Their latent space consists of manifolds that are entangled and highly non-linear. This makes it difficult to predict the results of modifications in the latent space on the output of the network. In this work, we investigate the latent space of shape networks, to try and develop techniques to obtain semantic deformations from latent editing operations. We consider two approaches: developing techniques to navigate complex entangles latent spaces, and developing less entangled and more interpretable representations, that can help in providing semantic editing operations.

Presenter: Ruben Vroegindeweij

Title: Depicting motion in a still image by spatio-temporal image fusion

Abstract: TBD

*Zoom meeting details upon request.

CGV Colloquium Friday September 11th

The CGV Colloquium will start online with a talk by Rana Hanocka (Tel Aviv University) on Friday, September 11 starting 16:00. The meeting is scheduled on Zoom, details are listed below.

Rana Hanocka is a 3rd-year PhD student in computer science at Tel Aviv University advised by Daniel Cohen-Or and Raja Giryes. At the intersection of Computer Graphics and Machine Learning, she is working on ways to use deep learning for manipulating, analyzing and understanding 3D shapes.

Title: Deep Learning on Single Shapes

Abstract:

One of the many difficulties in 3D deep learning is a lack of large, clean and labeled datasets. Acquiring and labeling large amounts of 3D data is not only cumbersome, but also requires using fundamental geometry processing pipelines that are not robust to geometry in the wild. Moreover, even high-quality 3D mesh models for similar shapes are extremely inconsistent with respect to triangulation and water-tightness, for example. On the other hand, training neural networks on single images has demonstrated surprisingly superb performance on a variety of different tasks. In this talk, I will present two recent works which propose training deep networks on a single shape. In Point2Mesh [SIGGRAPH 2020], we leverage the inductive bias of convolutional neural networks to learn a self-prior for surface reconstruction. We iteratively deform an initial mesh to “shrink-wrap” the input point cloud, resulting in a watertight mesh reconstruction. The weight-sharing property of CNNs models recurring and correlated structures within a single shape, and inherently removes noise and outliers. In Deep Geometric Texture Synthesis [SIGGRAPH 2020], we train a hierarchical GAN to learn to model the local geometric textures of a single shape. Our network displaces mesh vertices in any direction (i.e., in the normal and tangential direction), enabling synthesis of geometric textures, which cannot be expressed by a simple 2D displacement map.

*Zoom meeting details upon request.

CG Colloquium Thursday, March 12th

You are cordially invited to attend our CG Colloquium on Thursday, March 12th, 2020, 15:45-17:45h, Lecture Hall D@ta (Building 36)

The program features the following two speakers:

Prerak Mody

Title: 3D Human Pose Estimation Using a Top-view Depth Camera

Abstract: Delirium is a cause of concern within the health industry due to many postsurgery patients succumbing to this mental disease which disturbs their path to a full recovery. To understand and detect the onset of delirium within hospital ICU rooms, a depth camera (Microsoft Kinect) is attached to the ceiling. This depth data preserves privacy but also provides an opportunity to analyze the interactions taking place between the various stakeholders such as patient, hospital staff and visiting family. This project is being done at Philips Research, Eindhoven where my task is to extract the 3D human pose of individuals in the rooms. To this end, I extract the 3D point cloud data and run a supervised learning technique (i.e. 3D Convolutional Neural Network) to extract human pose. Having established a baseline, I am now investigating unsupervised and semi-supervised techniques to reduce the data and data annotation requirements respectively.

Thomas Saulou

Title: Photoshop for dummies : Energy-based image modification for photography composition

Abstract: Cameras have almost reached the limits in terms of hardware and optics, computational methods are now the major way to improve a photograph. However, too few tools are developed to enhance image composition. In this project, we introduce new methods based on photography rules to help photographers to modify the picture’s composition. We present a general approach to image deformation based on the energy, and applications of this approach to the problems of photography composition. Our method is inspired from works found in the prior art. The key advantage of our operator is the content-aware deformation function, which optimizes the location of the pixels modification. The operator has been developed to change lines composition in photographs.

CG Colloquium Thursday, February 27th

You are cordially invited to attend our CG Colloquium on Thursday, February 27th, 2020, 15:45-17:45h, room Kubus (building 26).

Johan van de Koppel, NIOZ

Title: Do self-organized patterns provide a solid basis for building infinite natural landscapes?

Abstract: Natural landscapes, when unaffected by human interference, are often characterized by a bewildering array of repeating natural patterns. Even in landscapes that are as good as flat, as are many dutch ecosystems, regular spatial patterns are found at many scales because of the interactions of organisms with physical forces, a process called spatial self-organization. In this talk, I will discuss how, by combining patterns at multiple spatial scales and including a multitude of processes, we can create visualisations of natural systems that are – in theory – infinite in size, without resorting to noise as the predominant factor creating heterogeneity spatial patterns occurring in the landscape, the ground and in the natural community (i.e. vegetation), we can build CGI representations that are theoretically infinite in scale. I will describe the origin of the patterns, and demonstrate how to build a visualization using patterns at different scales, using intertidal ecosystems as an example.

CG Colloquium Thursday, February 20th

You are cordially invited to attend our CG Colloquium on Thursday, February 20th, 2020, 15:45-17:45h, room Kubus (building 26).

Jim Whitehead

Title: Procedural Generation Using Linear Constraints

Abstract: Procedural content generation problems sometimes involve systems of linear equality and inequality constraints among the elements being generated. Cassowary is a constraint solver designed for solving these types of constraints at interactive speeds. It is widely used in user interface toolkits, and multiple implementations are available. To date, Cassowary has not been used in procedural content generation. This talk presents Cassowary and gives examples of its use in procedural content generation.

CG Colloquium Thursday, February 13th

You are cordially invited to attend our CG Colloquium on Thursday, February 13th, 2020, 15:45-17:45h, room Wit licht (building 26).

The program features the following two speakers:

Priyanka Bhaskar

Title: Narcolepsy – Exploratory Analysis

Abstract: Narcolepsy is a chronic neurological condition that results from the dysregulation of the sleep-wake cycle occurring in an early stage, specifically in adolescence. Although Narcolepsy occurs in an early stage there is a delay in diagnosis due to multiple reasons such as lack of symptom recognition, misdiagnoses as some of the symptoms overlap with other disorders like epilepsy, depression, insomnia to name a few. As a result of delayed diagnosis, the overall quality of life of a patient is affected.
The symptom pentad of narcolepsy includes excessive daytime sleepiness, cataplexy, hypnagogic hallucinations, sleep paralysis and disturbed nocturnal sleep. However, the symptoms related to narcolepsy are not limited to the pentad and cover a broad range covering symptoms not directly related to sleep, like, increase in weight, binge eating, anxiety, agitation. In order to reduce the delay in diagnosis, apart from having and creating an awareness about narcolepsy and it’s symptoms, it is important to look beyond the symptom pentad and understand the occurrences and interrelationships between these symptoms. Hence, this project aims at performing an exploratory analysis by looking into 20 symptoms, to understand symptom occurrences, the associations between symptoms to
name a few.

Joost Wooning

Title: Using Augmented Reality to assist in Medical Interventions

Abstract: Augmented Reality (AR) is an upcoming technique which can be used to display images overlaid on the normal vision of the user, several systems are on the market, in this project a Microsoft HoloLens will be used. A possible use of this technology is during medical interventions, allowing a surgeon to see displayed images inside a patient during surgery. If used as guidance a high accuracy of displayed projections is required to prevent damaging tissue. Previous work uses external tracking systems to enable accurate projections, using these systems is however not always possible, therefore this project will only use the hardware on a HoloLens. Previous research has shown that the spatial tracking of the HoloLens is not accurate, markers will, therefore, be used to track positions. The goal is to display a projection based on a CT image of a patient in the corresponding position inside the patient and report how accurate this projection is.

CG Colloquium Thursday January 16th

You are cordially invited to attend our CG Colloquium on Thursday, January 16th, 2020, 15:45-17:45h, EWI-Lecture Hall D@ta.

The program features the following two speakers:

Nouri Khalass

Title: Visualizing Axial-Symmetrical Nebula

Abstract: Nebulas are both interesting astrophysical phenomena as well as one of natures most beautiful sights to behold. Their astrophysical relevance comes from the fact that they can be the birthplace of new stars but also the remains of older stars. On top of that, nebulas have appeared countless times when depicting “space” in the entertainment industry. There exists a class of nebula that have a pronounced symmetry in their appearance. The goal is to create a tool that allows the user to model these kinds of nebula. By exploiting this symmetry the user does not have to model the complete nebula, but only a section of the nebula and from this a complete nebula can be synthesized.

Marie Kegeleers

Title: Interactive Story Authoring in Augmented Reality

Abstract: Augmented reality (AR) is a relatively new kind of technology that can be used as a tool to facilitate tasks that benefit from real time interaction and 3D visualisation in a real-world environment. An example of these tasks is story authoring. A story authoring application in AR allows the author to easily visualise story mechanics like plot points, as well as the authored scene with characters and props. Because AR integrates virtual elements with the real world, it offers more direct interaction with story elements which can provide a more intuitive experience compared to a PC application. This project introduces a new interface for AR in a tabletop environment using a head mounted device, that aims to facilitate the story authoring process through its different approach to interaction and visualisation. The interface provides a new way of interaction with virtual elements by combining both physical markers and hand gesture input. The tabletop environment is used to visualise the story authoring elements dynamically, in 3D, to avoid cluttering and enable adaptability. The goal of this combination of interaction and visualisation concepts is to provide a more intuitive story authoring experience.

CG Colloquium Thursday May 16th

You are cordially invited to attend our CG Colloquium on Thursday, May 16, 2019, 15:45-17:45h, at Pulse-Technology.

The program features the following two speakers:

Gijs Reichert

Title: Improving video analysis for the Olympic dinghy class sailing coaches and athletes

Abstract: Nowadays technology and data analytics are becoming increasingly intertwined with sports. The data can help assess performance during training and in competitive settings. However, for most sailors in the Olympic dinghy class it holds that the use of sensors during training is not standard practice and not allowed during races. To record and review their performance the coaches and athletes make use of video, shot (often by holding a camera in hand) from the a coach boat. The goal is to improve the capture, processing and analysis of videos used to train sailing athletes in the dinghy class. More specifically, the stabilization and quality of the footage taken from the coach boat that follows athletes will be improved. Next to this, of the entire recorded session only the interesting parts for the coach and athletes will be highlighted and extracted. To achieve this segmentation into clips object tracking will be used as well as methods to detect sailing manoeuvres. From these video clips the heel angle, position in boat and rudder movements can be extracted or enhanced to support the coaches in their assessment. The proposed approach should lead to a time-saving method to extract more information from the videos previously possible to support the training of the dutch Olympic sailing team.

Levi van Aanholt

Title: Semantic Tile Solving for Procedural Generation of Architectural Spaces

Abstract: We present an offline tile solving technique to generate complex architectural spaces. The generation is controlled by the notion of an architectural profile. An architectural profile allows to declaritively control spatial creation. It consists of semantic tiles representing building elements and declarative rules to control the generation according to architectural relations. This problem is converted to an ASP logic program and subsequently solved by an ASP solver. Our technique results in a declarative architecture generator that unifies procedurally generating building exterior, building interior and traversal in one system.

CG Colloquium Thursday May 2nd

You are cordially invited to attend our CG Colloquium on Thursday, May 2nd, 2019, 15:45-17:45h, at Pulse Hall 4

The program features the following three speakers:

Huinan Jiang

Title: Player model analysis for adaptive content delivery in an educational game

Abstract: Player in games has been a useful way for understanding player motivation, style and preferences, assisting us to predict their behaviour, adapt the and improve user experience. For serious games, learning is also an factor to consider. Most work done in this area is about adaptive difficulty. While in our educational game Squla, students are exposed to multiple game types, it’s interesting to know how their behavioural data could their game type preference, and whether custom game type delivery could improve students’ engagement and learning. Therefore we present a player analysis method game adaptation, and experiment with target student groups to look into the impact.

Antony Löbker

Title: Automatic reconstruction of real-world buildings using open data

Abstract: Using publicly available data , such as Google streetview panoramas and OpenStreetMap building, it is possible to reconstruct real-world buildings. While these are not highly detailed, they will give a general impression of a. The proposed method can quickly reconstruction an entire neighbourhood in major urban centers around the world.

Julie Hongping Feng

Title: Building a 4D MRI blood flow statistical model

Abstract: Cardiovascular (CVDs) are the global number one cause of death. Current of CVDs is mainly based on functions and morphology of cardiovascular, and 4D MRI blood flow data has become a new and powerful source of, which allows both anatomical and functional analysis in a single. However, it is still difficult to efficiently utilize this data source diagnose CVDs. One of the main reasons is the lack of standard analysis and understanding of this rather new imaging modality. The method to understand the healthy behaviour and its variation is to build an atlas. Current blood flow atlas do not support 3D velocity vector field full information. We attempt to build a 4D MRI blood flow atlas similar to statistical shape and appearance models by applying Principal Component Analysis (PCA) to get the mean and main variations of the blood flow fields. The of this atlas would help physicians understand the 4D MRI data and assist them to identify whether the data is abnormal or not. In order to more efficient representations of variations for vector fields, we also tested Complex PCA and Quaternion PCA, besides traditional Real PCA. Finally, will visualize the results such that they can be interpretable.

CG Colloquium Thursday April 18th

You are cordially invited to attend our CG Colloquium on Thursday, April 18, 2019, 15:45-17:45h, at Pulse Hall 4 .

The program features the following three speakers:

Remi van der Laan

Title: Enhancing Compression of the Sparse Voxel Directed Acyclic Graph

Abstract: Rendering massive scenes in real time represented as voxels has emerged as an attractive alternative compared to the traditional rendering pipeline. This has been achieved through the development of data structures that can efficiently store the scene data while also being inexpensive to traverse. The Sparse Voxel Directed Acyclic Graph is such a data structure, which losslessly compresses geometry by exploiting the spatial coherence in the scene. We attempt enhance the effectiveness of this compression through modifications to the construction of the graph and investigate possibility of applying lossy compression techniques.

Hao Ming Ye

Title: not available at the moment

Abstract: not available at the moment

Ruben Wiersma

Title: Graph Convolutional Networks for Learning on Point Clouds

Abstract: In the past decade, Convolutional Neural Networks (CNNs) have achieved incredible results. A recent development in the deep learning community is the attempt to generalise the advantages of CNNs from a Euclidean domain to non-Euclidean domains, like graphs and manifolds. Some examples of graph data are social networks, regulatory networks, functional networks, and 3D shapes. We attempt to get an understanding of the methods currently available and aim to improve on the current methodology for learning on point clouds. A tentative conclusion is that the conceptually simple Graph Convolutional Network by Kipf and Welling could be improved for manifolds through the incorporation of the connection Laplacian from vector diffusion maps.