Projects Proposed Projects

פרויקט זה עוסק בפיתוח מערכת לזיהוי תנועות יד באמצעות חיישן מכ"ם . מטרת הפרויקט היא לאסוף ולעבד נתוני מכ"ם, להפעיל ולהבין את הדוגמאות הקיימות של , ולאחר מכן לפתח אלגוריתם עצמאי המבוסס על טכניקות למידת מכונה לזיהוי מחוות יד. אחת המשימות המוצעות היא זיהוי הספרות 0 עד 9 בשפת הסימנים הישראלית, כולל מנגנון למחיקת טעויות. בשלב הסופי יוטען האלגוריתם לכרטיס המכ"ם ויודגם בפעולה

Denis Dikarov

Yaron Honen

Automotive radar is a key sensing modality for autonomous vehicles, but collecting and annotating diverse radar data is expensive. This project investigates whether simulation-based data augmentation can improve radar object detection on real-world data. A key challenge is that generating new radar scenarios requires first creating realistic changes in the camera inputs, such as adding or removing vehicles, while keeping these edits consistent across multiple camera views and time. The project builds on RadarGen, a generative model that synthesizes automotive radar point clouds from multi-view camera observations. The student will explore modern generative image, video, and multi-view editing methods and develop a pipeline for controllable vehicle insertion and removal with consistent geometry and automatically updated annotations. The edited scenes will then be passed through RadarGen to generate corresponding radar measurements. These synthesized measurements, together with their automatically generated labels, will be used to train a radar object detector. The teaser illustrates the overall pipeline from multi-view imagery, through RadarGen, to radar-based detection. The main research question is whether these generatively augmented multi-view scenes, and the radar data synthesized from them, can improve detection performance on real, unseen radar data.

Tomer Boreda

Yaron Honen

גילוי רחפנים זו בעיה חשובה בתחום הביטחון, לשדות תעופה, וליישומים אחרים.

הרחפנים יכולים להיות קטנים ולכן הגילוי שלהם קשה, במיוחד ממרחק גדול.

במסגרת הפרויקט, נפתח מערכת לגילוי רחפנים הכוללת חומרה חישובית, מצלמות, ואלגוריתמים.

נפתח אלגוריתמים לגילוי בפריים בודד ובווידאו על סמך מאפיינים ייחודיים של הרחפנים, ביצועי המערכת ינותחו על סמך אמינות הזיהוי הנדרשת בתרחיש אמת מול קצב גילויי שווא נמוך ככל האפשר.

נעבוד עם Datasets בווידאו של רחפנים עם Ground truth

Tal Nir

Yaron Honen

הפרויקט מתאים ל- 3-6 סטודנטים בעלי רקע באחד או יותר מהתחומים הבאים בראייה ממוחשבת, עיבוד תמונות, ולמידה עמוקה בהקשרים ויזואליים. 

Recent years have seen rising interest in facial reconstruction for both static and dynamic 3D scenes. Gaussian Splatting has emerged as a compelling representation, enabling real-time rendering and high-fidelity novel-view synthesis. At the same time, text-driven 3D facial editing has shown promise, but most existing methods primarily manipulate appearance parameters (e.g., per-Gaussian color, opacity, scale), leaving geometry largely unchanged and limiting expressiveness.

Eldad Matmon

Long-range optical flow and dense tracking are powerful tools for applications such as 3D reconstruction, animal behavior analysis, and motion analysis. Another promising application is video editing, where edits made to a single frame can be propagated through a video sequence using optical flow, simplifying the task to image-level editing. However, current optical flow systems often suffer from drift errors and struggle with long-term occlusions, resulting in editing “holes” and inconsistencies. The recent Dense Optical Tracking (DOT) method has achieved state-of-the-art performance in long-range optical flow by combining a sparse set of long-range tracks with iterative refinement, providing superior occlusion handling and improved long-term stability.  This project aims to address the limitations of existing optical flow systems by leveraging DOT’s accurate long-range correspondences and, potentially, refining the final edits using VDM-based methods. This approach promises to improve occlusion recovery, reduce drifting errors, and enhance the quality of propagated edits. Students involved in this project will engage with cutting-edge advancements in video processing and generative modeling, building a strong foundation for further research and gaining valuable theoretical and practical skills.

Assaf Singer

Establishing dense point correspondences in video has seen remarkable progress in recent years, expanding recently to long-range point tracking.  DINO-Tracker, we introduced a novel self-supervised method for long-range dense video tracking, harnessing DINO's powerful visual priors. Our approach combines test-time training on individual videos with the highly localized semantic features learned by a pre-trained DINO-ViT model, yielding state-of-the-art results, particularly in tracking across extended occlusions. Further details can be found on our project page. While DINO-Tracker shows promise, its optimization process—taking 1-2 hours per video—limits its practicality for real-time applications. This project aims to address this by exploring feed-forward, real-time solutions that enhance temporal consistency across frames and increase DINO's spatial resolution, enabling high-resolution tracking capabilities. This project offers students the opportunity to engage with cutting-edge advancements in video tracking and processing, providing a solid foundation for further research in this dynamic field. Participants will gain expertise in both theoretical and practical aspects of video tracking and representation learning.

Assaf Singer

We recently introduced a novel photo-realistic image restoration algorithm based on flow matching, a generalization of diffusion. For more details, visit our project page: https://pmrf-ml.github.io.

We would like to continue with this line of work by:
1. Designing our algorithm to trade distortion (e.g., MSE) with visual quality at test time (by tuning a test time hyper-parameter).
2. Considering recent and more advanced schemes of flow matching, potentially improving both the theoretical guarantees and the practical results of our method.
This project provides an opportunity for students to engage with cutting-edge developments in photo-realistic image restoration, paving their way for further research in this dynamic field. The project will emphasize learning, brainstorming, and developing new mathematical insights rather than extensive coding hours. Participants will gain expertise for conducting both theoretical and practical research in the endless field of image restoration

Guy Ohayon

Daniel Silver

As a continuation of our latest publication, we thought of an elegant, theoretically-based solution to assess the determinism & robustness of an image restoration algorithm via the diversity of its outputs. 

Guy Ohayon

In recent years, unmanned aerial vehicles (UAV) are exploited to capture images of large agricultural areas. For agricultural purposes, near infrared (NIR) or other spectral channels are acquired in addition to the RGB colors. These are then used for various tasks, especially semantic segmentation – the ability to distinguish between healthy plants, unwanted weeds, running or standing water, and other classes. We will use an existing semantic segmentation framework, test several network architectures, and employ pre-processing and post-processing algorithms for improved results. Our goal is to participate in an international agriculture segmentation challenge, and attempt to achieve high scores. 

Alon Zvirin

Yaron Honen

Instance segmentation is the task of detecting and masking objects in an image and distinguishing between instances of the same class. Accurate segmentation of leaves in plant images is important in many agricultural applications. These include early detection of water and heat stress, identification of biological infection, monitoring of plant growth, and prediction of harvest yields. Our purpose is to use an existing instance segmentation deep neural network, and integrate it with image processing tools for better performance. In this project we will improve upon previous results, participate in an international leaf segmentation challenge, and attempt to achieve best scores

Alon Zvirin

Yaron Honen

A few years ago, we showed for the first time that the molecular profile of cancer can be predicted by analysis of biopsy images, without using molecular assays. In other words, the shape of the tumor cells and the tissue architecture hold information that allows to accurately predict molecular expression, even though such molecules cannot be seen by humans by visual examination of biopsy images.  Based on these findings, in the past two years we established collaborations with several medical data hospitals in Israel and abroad and extended the scope of our research to different prediction tasks in breast cancer, lung cancer, and Leukemia. We collected and scanned high quality well annotated tens of thousands of histology slides, and extended our research team, consisting of data scientists, graduate and postdoctoral students and clinical collaborators and advisors. We have recently shown how to steer our technology into assisting personalized medicine procedures. Take a part in developing an AI-based framework for analysis of histology images for improving personalized oncology  -  Link to the article 

Gil Shamai