ISP Seminars

Previous Seminars

Recovery guarantees for linear and non-linear reconstruction from Walsh measurements
Laura Thesing (University of Cambridge)(Invited talk)
28 October 2020
In this talk we discuss recovery guarantees for the reconstruction of wavelet coefficients from Walsh measurements, which appear in fluorescence microscopy, lensless cameras and other analogue measurement advices with an “on-off” behaviour. We consider both linear and non-linear reconstruction methods. For the analysis we discuss the typical structure of signals under the wavelet transform and the properties of the change of basis matrix. The results offer a guideline on the choice of sampling pattern and an insight to the relationship between Walsh functions and wavelets.
Ressources :

(Online) Microsoft Teams -- Wednesday, 28 October 2020 at 14:00 (45 min.)

Re-Identification for Multi-Person Tracking
Vladimir Somers
22 July 2020
Continuous identification of sport players during a game is essential to produce personalised content and collect individual statistics. With images provided by a single camera view setup, player identification relies on two components: (1) player recognition, which can be performed only sporadically, when discriminative appearance features are available and (2) individual player tracking during the game using detections generated at each frame. Individual tracking is easy when the player is isolated on the court, but becomes difficult or impossible when there are occlusions and complex interactions with other players. For that reason, long-term tracking is generally implemented based on the temporal association of shorter players tracks, also called tracklets, which are obtained by non ambiguous association of detections from the same player. Numerous reliable and efficient solutions exist to generate these short tracklets. We will therefore focus our work on the temporal association of tracklets based on their underlying detections appearance. This requires estimating the affinity between pairs of tracklets, to identify the pairs that correspond to the same identity or to distinct ones, but also to embed those cues in a graph-based formalism to jointly optimize the tracks of multiple players. In this presentation, we will explain how to use state of the art CNN based models for person re-identification in order to address the question of visual affinity estimation between pairs of tracklets in a team sport multi-person tracking context.
Ressources :

(Online) Microsoft Teams -- Wednesday, 22 July 2020 at 14:00 (45 min.)

Synthetic corruption of images for anomaly detection using autoencoders
Anne-Sophie Collin
3 June 2020
Anomaly detection can be defined as the process of identifying rare items, events or observations that differ significantly from the majority of the data. In industrial vision, this problem can be addressed with an autoencoder trained to map an arbitrary image (with or without any defect) to a clean image (without any defect). In this approach, anomaly detection relies conventionally on the reconstruction residual or, alternatively, on the reconstruction uncertainty. The higher the reconstruction residual/uncertainty, the higher the likelihood of a region being defective.
Ressources :

(Online) Microsoft Teams -- Wednesday, 3 June 2020 at 14:00 (45 min.)

Joint use of semantic and geometric information for object localization using time-of-flight cameras
Victor Joos de ter Beerst
18 May 2020
Time-of-Flight (ToF) cameras are great for the mapping of close-up objects, but when applied to the analysis of large scenes, the task becomes more challenging due to low resolution, distance errors, and occlusion among others.
Ressources :

(Online) Microsoft Teams -- Monday, 18 May 2020 at 11:00 (45 min.)

Using Approximate Computing to Improve the Efficiency of LSTM Neural Networks
S. Abolfazl Ghasemzadeh
4 March 2020
As the growing field of Artificial Neural Networks, Recurrent Neural Networks are often used for sequence-related applications. Long Short-Term Memory (LSTM) neural networks are improved and widely used versions of Recurrent Neural Networks. To achieve high accuracy, researchers always build large-scale LSTM networks which are time-consuming and power-consuming.
Ressources :

Shannon seminar Room (Place du Levant 3, Maxwell Building, 1st floor) -- Wednesday, 4 March 2020 at 11:00 (45 min.)

Topological and molecular characterization of pancreatic endothelial cells: a prerequisite for tissue bioprinting
Laura Glorieux (PhD student at De Duve, UCLouvain) (Invited talk)
11 February 2020
Developing therapies for pancreatic diseases is hampered by limited access to pancreatic tissue in vivo and engineered 3D tissue models have great potential for physiopathological and pharmaceutical research. An important and limiting step in tissue engineering is tissue vascularization. This project is part of the Pan3DP consortium that aims to bioprint murine embryonic pancreatic tissue units. The first step of this project is to collect information on the 3D architecture and transcriptomic profile of a developing pancreas, then to bioprint pancreatic tissue units based on a digital 3D pancreatic atlas, and finally to mature these units into functional tissue under custom-optimized in vitro conditions. Here, we focus on the spatial characterization of the endothelial compartment and on the integration of the vascular network in bioprinted construct.
Ressources :

Shannon seminar Room (Place du Levant 3, Maxwell Building, 1st floor) -- Tuesday, 11 February 2020 at 13:00 (45 min.)

Data fitting on manifolds: applications, challenges and solutions
Pierre-Yves Gousenbourger
11 December 2019
Storm trajectories prediction, birds migrations follow-up, rigid rotations of 3D objects, wind field estimation, model order reduction of superlarge parameter-dependent dynamical systems, MRI 3D body volumes reconstruction... All these applications have two things in common: first, they have a geometrical data-structure, i.e., the data lives on a (generally) Riemannian manifold; second, they can benefit of parameter(s)-dependent fitting methods somewhere in the process. If data fitting is a basic problem in the Euclidean space (where natural cubic splines and thin plates splines are the superstars in the domain), it become more intricate when data structure constrains the problem. This talk is an opportunity to present you an efficient, ready-to-use algorithm for data fitting on manifolds based on Bezier curves, applied to some of the aforementioned applications.
Ressources :

Euler seminar room (Room A.002, Euler Building, Avenue Georges Lemaître 4-6) -- Wednesday, 11 December 2019 at 11:00 (45 min.)

Tone-mapping for X-ray images
Tahani Madmad
6 November 2019
Inspecting X-ray images is an essential aspect for medical diagnosis and for nondestructive control of manufactured objects in the industrial field. However, X-ray images are characterized by a high dynamic range and a low contrast. Due to those characteristics, and the limitation of the human visual system, important aspects such as nodules, bones fractures, gas inclusions or other kind of defects and anomalies are difficult to identify for the human eyes. Through my presentation, I will first introduce the challenges of X-ray images visualization and give an insight into the state of the art. At a later stage, I will present the fusion framework I adopted to address this issue and then focus on an instance of this vision that led to a tone-mapping operator based on a bilateral histogram equalization. Finally, I will discuss the options I am considering to overcome the limitations of this tone-mapping algorithm and achieve the genericity needed to visualize highly contrasted X-ray images regardless of their content.

Shannon Seminar Room (Place du Levant 3, Maxwell Building, 1st floor) -- Wednesday, 6 November 2019 at 11:00 (45 min.)

Can we scale open-source development by removing the need for iconic and self-sacrificing leaders?" (slides)
Simon Carbonnelle
2 October 2019
Today's most popular end-user applications are owned and controlled by private companies (e.g. Facebook, Youtube, Reddit, Gmail, Github,...).
Ressources :

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Wednesday, 2 October 2019 at 11:00 (45 min.)

An outlook on deep reinforcement learning
(invited talk) Dr Vincent François-Lavet (McGill University, Canada & Mila)
6 September 2019
Deep reinforcement learning is the combination of reinforcement learning (RL) and deep learning. This field of research has been able to solve a wide range of complex decision-making tasks that were previously out of reach for a machine. In this talk, I will provide an introduction to deep reinforcement learning models, algorithms and techniques. I will particularly focus on the aspects related to generalization of the learned policy to slightly different situations than the ones encountered during training.
Ressources :

Euler building (room A.002) -- Friday, 6 September 2019 at 11:00 (45 min.)