Speaker: Oran Gafni

M.Sc. student under the supervision of Prof. Lior Wolf

Wednesday, February 27^th 2019 at 15:30

Room 011, Kitot Bldg., Faculty of Engineering

Live Semantic Face Editing in Video using Deep Adversarial Autoencoders

Abstract

We propose a method for face editing in video that enables live face effects at high frame rates. Two applications are considered (i) replacing the face with a similar face that is not recognizable as the same identity, and (ii) modifying parts of the face. These applications require maintaining the pose, the apparent illumination, and the expression of the face in the input frames while making natural-looking modifications according to the desired task.

We achieve this by a novel feed forward encoder-decoder architecture that is conditioned on the target high-level features of a single image. The network is global, in the sense that it does not need to be retrained for a given video or based on the desired outcome, and it creates naturally looking sequences with little distortions.

Speaker: Matan Shoef

M.Sc. student under the supervision of Prof. Daniel Cohen-Or

Monday, February 25^th 2019 at 15:30

Room 011, Kitot Bldg., Faculty of Engineering

PointWise: An Unsupervised Point-wise Feature Learning Neural Network

Abstract

We present a novel approach to learning a point-wise, meaningful embedding for point-clouds in an unsupervised manner, through the use of neural-networks.

The domain of point-cloud processing via neural-networks is rapidly evolving, with novel architectures and applications frequently emerging. Within this field of research, the availability and plethora of unlabeled point-clouds as well as their possible applications make finding ways of characterizing this type of data appealing. Though significant advancement was achieved in the realm of unsupervised learning, its adaptation to the point-cloud representation is not trivial.

Previous research focuses on the embedding of entire point-clouds representing an object in a meaningful manner. We present a deep learning framework to learn point-wise description from a set of shapes without supervision.

Our approach leverages self-supervision to define a relevant loss function to learn rich per-point features. We train a neural-network with objectives based on context derived directly from the raw data, with no added annotation. We use local structures of point-clouds to incorporate geometric information into each point's latent representation. In addition to using local geometric information, we encourage adjacent points to have similar representations and vice-versa, creating a smoother, more descriptive representation.

We demonstrate the ability of our method to capture meaningful point-wise features through three applications. By clustering the learned embedding space, we perform unsupervised part-segmentation on point clouds. By calculating Euclidean distance in the latent space we derive semantic point-analogies. Finally, by retrieving nearest-neighbors in our learned latent space we present meaningful point-correspondence within and among point-clouds.

School of Mechanical Engineering Seminar
Thursday, March 6, 2019 at 14:00
Wolfson Building of Mechanical Engineering, Room 206

Finding the Limits of Submerged Jet Heat Transfer

Barak Kashi

PhD student of Herman Haustein

Impinging jets offer much higher transport rates than other canonical flows and therefore have many industrial applications. Specifically, for cooling of electronic components they are often employed in arrays of micro-jets for improved uniformity. Examination of the literature shows that prior to dealing with arrays, key understandings of single jet heat transfer are still lacking. Within this study, these gaps are filled by new relations which are developed and validated against numerical simulations over a wide range of working conditions.

While traditional macro-scale heat transfer (HT) correlations predict continuous increase in HT during down-scaling, it is here found that this trend reverses at the micro-scale due to viscous dissipation even under low-viscosity laminar flow conditions. Thus, a maximum exists for effective cooling, beyond which any further down-scaling decreases the cooling rate. Furthermore, It is shown that when dissipation becomes significant, the effective HT is no longer independent of the imposed heat flux and the heating-cooling symmetry is broken.

Next, dependencies of HT on geometrical parameters such as nozzle length and nozzle-to-wall distance are resolved,  showing that the velocity profile arriving at the wall uniquely determines the stagnation-point HT. First examining short-nozzle flow reveals three distinct flow regimes, associated with flow separation at the inlet and its reattachment. The corresponding non-monotonous dependence of HT on nozzle length is then correlated for short nozzle-to-wall distances. Larger nozzle-to-wall distances are here dealt with by deriving an approximate solution to the axial momentum equation, which gives a good description of the jet core for a wide range of flight distances and for all issuing profile. This approximate solution, not only revealed the true form of the jet’s potential core, but also led to the identification of a new concept – the “boundary core” delimiting the jet’s interaction with its surroundings.

Finally, a mechanistic model is developed which provides the HT distribution under the jet and its relation to the arriving jet profile. Therein, apparent contradiction in observed stagnation-point HT trends is resolved, while the non-monotonous radial distribution under quasi-uniform jets is explained.

The contribution of the present work is two-fold: i) identifying and quantifying the inherent lower bound of beneficial down-scaling of jet impingement systems­ – thus curbing the on-going trend of jet-array miniaturization; ii) the relations developed here significantly extend jet impingement theory and enable finding the optimal HT regardless of scale.