refaev Goto Github PK

Email: [email protected]

LinkedIn: Refael Vivanti

I do R&D in Reinforcement Learning, Deep Learning, SLAM and Computer Vision. Currently I'm at Meta and. HUJI.

TOP GITHUB PROJECTS:

two-armed-bandit - some reinforcement learning research on what happends when there's a different reward variance in some (two) policies.

robots routing - two robots navigate between obstacles while avoiding each other.

intercepetor challenge - a challange in Rafael which we solved using deep reinforcement learning.

Sadly, most of my work can't be shared; Rafael and Meta are closed-source, and my PhD was in HUJI-internal repos.

2021 – current: lecturer at the Hebrew University, Jerusalem Israel. Course 67604 Vision Aided Navigation.

In this course we try to give the students a hands-on experience in implementing a Simultaneous Localization And Mapping (SLAM) system for vehicle navigation using a stereo pair. Syllabus.

2021 - current: Computer vision Engeneer, Meta Inc.

I do cloud localization for AR/VR devices. Our main challenge is to relocalize an AR or VR device using large-scale maps with low compute and delay.

2017 – 2021: Researcher, Rafael ltd

• Deep Reinforcement Learning for obstacle avoidance. I took a DRL algorithms such as PPO and A2C, added my algorithmic improvements and tuning, and trained an agent for vehicle autonomous navigation and driving using only images as input in an end-to-end manner. I integrated it into an actual driving robot.

Simulation video

real video

lecture

• During this research, I spotted a new theoretical issue I call 'the manipulative consultant problem', and by solving it I boosted the results dramatically. paper code

• Deep Reinforcement Learning for image matching – I've implemented a search-guiding agent for image retrieval task. It searches the infrastructure like a human may do, and cut ~50% of searching time.

• Obstacle avoidance in geometric methods such as Stereo-based depth reconstruction, deep learning for freeway segmentation and RADAR tracking.

• Epipole correction from multiple views.

• Image segmentation of aerial images using VAE architecture, and thermal sensor simulation using GAN architecture.

All of the work was done in Python, using Keras, Tensorflow or Pytorch.

2010 – 2016 Senior Computer Vision Algorithms designer, Rafael ltd

My main area was Simultaneous Localization And Mapping (SLAM) and classic and deep machine learning for various applications:

• Automatic mapping and 3d estimation of large scale areas, such as whole countries.
• Automatic electric-pole detection from aerial images using Classic Machine learning.
• Automatic vehicle and pedestrian detection from aerial images using machine and deep learning.
• 3D estimation and depth reconstruction from aerial images
• Accurate computer vision based navigation correction
• Active-Learning system for semi-automatic target classification, using a machine-learning boosting method.
• Post mortem track reconstruction from video input.
• Stable out-of-image tracking.
• Time to arrival estimation using computer vision.

Most of the work was done in C++ and Matlab

2005 – 2010 Computer Vision engineer, Rafael ltd Implementing real-time SLAM systems:

• Video tracking
• Outlier removal
• Bundle adjustment
• Soft Real-Time optimization
• Image processing algorithmic optimization
• GUI implementation for aerial mapping. The work was done in C++ and C# with C interfaces

Note: Due to security classification issues, much of the work I done cannot be detailed, and most of the code for the described project is classified.

2012-2018: PhD in Computer Science at the Hebrew University, Jerusalem, Israel

My PhD was in Medical Image Processing in the Hebrew University, under the supervision of Prof. Leo Joskowicz.

My main subject was using Machine Learning Algorithms like Deep Neural Networks for the detection and accurate segmentation and tracking of abdominal and pulmonary tumors, in longitudinal CT studies.

I also used deformable tissue registration, and video tracking of deformable tissues. Work was done in Matlab and C++

2005-2008: MsC Cum Laude in Computer Science at the Hebrew University, Jerusalem, Israel Avg: 92

Thesis: Facing the challenge of creating an Augmented Reality solution for laparoscopic nephrectomies procedures (partial renal resection), I used novel computer vision algorithms for the 3D segmentation on kidney anatomies in preoperative CT scans, as a pre-step for AR presentation to the surgeon. thesis

2001-2004: B.Sc Cum Laude in Computer Science, bioinformatics and math at the Hebrew University, Jerusalem, Israel Avg: 93.1

Final project: Protein secondary structure predictions using SVM and boosting methods.

• Vivanti, R. Talya D. SB, Shlomo C, Orna C, Adaptive Symmetric Reward Noising for Reinforcement Learning, arxiv.org/abs/1905.10144

• Vivanti, R., Szeskin, A., Lev-Cohain, N., Sosna, J. and Joskowicz, L., 2017. Automatic detection of new tumors and tumor burden evaluation in longitudinal liver CT scan studies. International journal of computer assisted radiology and surgery, 12(11), pp.1945-1957.

• Vivanti, R., Ephrat, A., Joskowicz, L., Karaaslan, O.A., Lev-Cohain, N. and Sosna, J., 2015. Automatic liver tumor segmentation in followup CT studies using convolutional neural networks. In Proc. Patch-Based Methods in Medical Image Processing Workshop (Vol. 2).

• Vivanti, R., Ephrat, A., Joskowicz, L., Lev-Cohain, N., Karaaslan, O.A. and Sosna, J., 2015, October. Automatic liver tumor segmentation in follow-up CT scans: preliminary method and results. In International workshop on patch-based techniques in medical imaging (pp. 54-61). Springer, Cham.

• Vivanti, R., Joskowicz, L., Karaaslan, O.A. and Sosna, J., 2015. Automatic lung tumor segmentation with leaks removal in follow-up CT studies. International journal of computer assisted radiology and surgery, 10(9), pp.1505-1514.

• Faria, C., Sadowsky, O., Bicho, E., Ferrigno, G., Joskowicz, L., Shoham, M., Vivanti, R. and De Momi, E., 2014. Validation of a stereo camera system to quantify brain deformation due to breathing and pulsatility. Medical physics, 41(11), p.113502.

• Shlomo Shenzis, Moshe Samson, Refael Vivanti, Leo Joskowicz, Naama Lev Cohain and Jacob Sosna. 2015. 3D segmentation using perceptual computing. Interactive Medical Image Computation Workshop, MICCAI 2015, Munich, Germany, Oct 9, 2015.

I started and admining some large technical communities in Israel with more than 5K local experts: Computer Vision Israel Reinforcement Learning Israel Bioinformatics Israel

I'm also admining jobsJLM to help promote the hi-teck industry in my city Jeruslaem.

reinforcement-learning-is-full-of-manipulative-consultants