Manoosh Samiei

I am an AI Researcher at Mila, supervised by Aaron Courville. Previously I worked for two years as a computer vision researcher and 3D reconstruction engineer at Algolux, a self-driving car software startup, and Magicplan, Sensopia Inc., an augmented reality mobile app which maps indoor environments.

I did my masters at McGill in Electrical and Computer Engineering, where I was advised by James J. Clark. My master's research was on modeling visual attention and distraction, which is closely related to the saliency prediction, in visual search tasks using deep learning and eye tracking data.

I did my undergraduate at Shahid Beheshti University in Electrical Engineering with a focus on telecommunications and signal processing. My undergraduate research was focused on end-to-end training approach for lane following task (behavioral clonning) in autonomous vehicles using convolutional neural networks. During my undergraduate, I worked on multiple robotic projects and learned about electronic circuits troubleshooting and assembly.

My research interests lie at the intersection of human vision, robotics, computer vision and machine learning. Besides AI and programming, I also enjoy hardware assembly and electronic circuitry.

Email  /  LinkedIn  /  Google Scholar  /  Twitter  /  Github

15/09/2023 I am joining Mila soon as an AI researcher!

09/08/2022 Moving to Toronto, Ontario on 1st September!

01/08/2021 I am officially graduated from McGill with a Master of Science in Electrical Engineering!

16/08/2021 I am starting as a computer vision researcher at Algolux!

We present two approaches. Our first method uses a two-stream encoder-decoder network to predict fixation density maps of human observers in visual search. Our second method predicts the segmentation of distractor and target objects during search using a Mask-RCNN segmentation network. We use COCO-Search18 dataset to train/finetune and evaluate our models.

DeepGaze2 extracts high-level features in images using VGG19 convolutional neural network pretrained for object recognition task. DeepGaze II is trained using a log-likelihood learning framework, and aims to predict where humans look while free-viewing a set of images.

We implmented two papers that formulate object localization as a dynamic Markov decision process based on deep reinforcement learning. We compare two different action settings for this MDP: a hierarchical method and a dynamic method.

We replicated the results of the paper “CNN2: Viewpoint Generalization via a Binocular Vision” for two datasets SmallNORB and ModelNet2D.

Source code and style from Jon Barron's website