Souhaiel BenSalem

I’m a Deep Learning Research Engineer at Parrot Drones in Paris, specializing in Vision for Robotics. My work involves developing depth estimation algorithms for obstacle avoidance, theoretical and empirical studies on neural networks, and exploring domain adaptation and generalization to improve model robustness.

Prior to that, I was a Research Intern within Inria Paris' and valeo.ai's joint team Astra-Vision, where I worked on 3D scene understanding via transferring the knowledge of 2D VLMs / VFMs under the supervision of Raould de Charette, Renaud Marlet, Gilles Puy and Alexandre Boulch.

Before that, I also spent time at Inria Lyon's CHROMA as a Research Intern, working with Jilles Dibangoye> on establishing theoretical foundations to near-optimally solving Dec-POMDPs under hierarchical information sharing.

I have received an MSc. degree in Mathematics, Vision, Learning from ENS Paris Saclay, an MSC. degree in Mobile Autonomous Robotic Systems from Grenoble-INP and an Engineering degree in Robotics and Mechatronics from the the university of Carthage.

Email  /  Resume  /  Github  /  Linkedin  /  x

Research Interests

My research interests lie at the intersection of visual cognition, learning theory (self-supervised and unsupervised paradigms), and representation learning. I am particularly interested in the role of vision within the action-perception loop, where it drives learning and decision-making processes in complex, visually-rich environments and how to develop (predictive) models to learn abstract, hierarchical representations that enable robust and intuitive interactions of agents / robots in these environments.

Academic Projects

The following are some selected academic and personal projects. More are available on my GitHub.

Evaluated the VICReg method on smaller datasets, pretrained a visual backbone, and extended the evaluation of an ImageNet-pretrained model to new downstream tasks.

Tackled the DIV2K dataset using SRGAN, ESRGAN, and SRResNet methods. Proposed a novel perceptual loss function and evaluated performances for different loss functions.

Implemented and compared various reinforcement learning algorithms (PPO, A2C, DQN) to solve the Flappy Bird game. Achieved optimal performance using Dueling Double DQN, outperforming baseline models with improved convergence and reward stability.

Attempted to recreate DXO's ClearView Plus technology. Studied, implemented, and adapted the image dehazing technique introduced by Meng et al.

Implemented and compared model-free reinforcement learning (Q-Learning, SARSA, Dyna-Q) and deep RL algorithms (DQN, REINFORCE, A2C) to solve the Maze Rider problem. Used reward shaping, efficient sampling, and state space reductions to improve convergence and performance.

Implemented the Exposure Fusion technique, eliminating the need for physically-based HDR assembly. Achieved image quality comparable to traditional tone mapping operators.

Designed and trained a U-Net model for advanced audio denoising tasks. Leveraged predicted noise magnitude spectrograms for noise separation and used inverse Short-Time Fourier Transform (STFT) for accurate original audio reconstruction.

Implemented a control strategy for a vehicle based on the Bicycle-Model, focusing on both lateral and longitudinal dynamics. Integrated Model Predictive Control (MPC) for automated steering and a PI controller for longitudinal speed tracking. Unified both strategies and validated performance using simulations with the Pure Pursuit algorithm.

Transfer learning for violence detection using a CNN-LSTM-based model.

Deployed the MNIST handwritten digit recognition model on the STM32F429 DISCOVERY board, showcasing deep learning on ARM-based microcontrollers.

Blog Posts

Miscellanea

I am passionate about reading (research papers included), reverse-engineering, (discussing) anthropology, and game theory.

Fun Fact: Me and Luke Skywalker share the same hometown! Tozeur, a city in southern Tunisia, where Star Wars: A New Hope brought Planet Tatooine to life, and where iconic scene where Luke stood gazing at the binary sunset was filmed. A scene that symbolizes the closing of one chapter and the dawning of another .. and of cource, a new hope.