Ozi Marom,2013

The project developed a stereo camera rig system to capture multiple synchronized views of growing vegetables and reconstruct their underground root structures in 3D. By applying multi-view geometry and shape-from-silhouette or photogrammetric techniques, the system produces accurate root models that enable continuous, non-destructive measurement of root growth, morphology, and spatial development over time.

Alex Gusin,2013

The project developed a system that automatically tags browser bookmarks by analyzing the content and metadata of each saved webpage. Using lightweight text extraction, topic modeling, and semantic classification, it assigns meaningful, consistent tags that reflect the page’s subject and intent. This enables faster retrieval, improved organization, and intelligent bookmark grouping without manual labeling. 

Ran Mishael,2018

The project built a unified model that performs image segmentation and super-resolution simultaneously, allowing each task to reinforce the other for superior visual quality and accuracy. By recovering fine-grained details while delineating object boundaries in a single framework, the system produces cleaner segmentations, sharper textures, and more reliable downstream insights. This joint approach reduces computational overhead, improves robustness in low-resolution or noisy imagery, and delivers clearer, more informative images for analysis and decision-making. 

Dan Basson,2019

The project explored curriculum learning for image classification, organizing the training data from simple to complex examples to guide the model’s learning. By structuring the learning progression, it improved training stability, accelerated convergence, and enhanced the classifier’s ability to generalize from challenging samples. This approach ultimately led to better performance than models trained on randomly ordered data. 

Yael Galifat,2020

The project investigated deep learning model optimization through selective inference, in which the system dynamically determined which inputs required complete model evaluation and which could be handled by lighter, approximate predictors. By avoiding unnecessary computation on easy or redundant cases, it reduced latency and resource usage without compromising accuracy. This strategy ultimately improved overall efficiency, enabling faster and more cost-effective deployment of deep learning models in real-time settings. 

Barak Katz,2020

The project focused on detecting human interactions in egocentric video, developing a model that analyzed first-person visual streams to identify when and how the camera wearer engaged with others. By combining motion cues, pose patterns, and spatial-temporal features, the system recognized interaction events such as conversations, handovers, and collaborative actions. This approach enabled more accurate interpretation of social behavior from wearable cameras and demonstrated improved performance over traditional frame-based methods. 

Inga Efrosman, Daniel Cohen,2021, report

The project develops an image-mosaicking system that uses SLAM techniques to stitch overlapping views into a high-resolution, coherent panorama. By estimating camera motion, refining feature correspondences, and maintaining a consistent global map, the system produces seamless mosaics even in large or visually complex environments. This approach offers greater robustness than traditional stitching methods and enables accurate scene reconstruction from unconstrained camera motion.  

Yaniv Goren,2021, report

Interjection Speech Recognition, as a voice-touch interface, focuses on detecting short, natural vocal bursts such as hmm, oh, ah, or hey and interpreting them as lightweight interaction commands. The project develops models that distinguish these brief, expressive sounds from regular speech, classify them into meaningful categories, and map them to interface actions like confirmation, hesitation, selection, or alerting. This enables fast, intuitive, hands-free interaction, particularly useful in scenarios where users cannot rely on traditional touch or full-sentence voice commands.  

Eyal Reis,2023

A vision-based system for recognizing control elements in vehicle dashboards, identifying buttons, knobs, indicators, and display regions from onboard camera images. By combining object detection, fine-grained classification, and layout analysis, the system interprets the dashboard’s structure and labels its functional components. This enables applications such as driver-assistance interfaces, automated documentation, and enhanced human–vehicle interaction through reliable understanding of in-cabin controls. 

Aharon Sahalulu,2023

A haze-aware object detection system that compensates for reduced visibility and contrast in hazy or foggy conditions. By integrating dehazing priors, visibility-adaptive feature extraction, and robustness-enhanced detection models, the system maintains reliable recognition of vehicles, pedestrians, and other critical objects even in degraded atmospheric environments. This improves detection accuracy, enhances safety, and supports consistent performance in real-world outdoor scenarios. 

Rom Hirsch, Yarom Swisa,2024

A classification system explicitly designed for dark or low-light images, incorporating illumination-aware preprocessing and models robust to noise, low contrast, and color distortion. By enhancing essential features while suppressing artifacts introduced by poor lighting, the system achieves reliable recognition where standard classifiers fail. This enables more accurate analysis of nighttime scenes, security footage, and other low-visibility imagery. 

Ben Kedarya, Yael Rikka,2024

An audio alarm detection system tailored for hearing-impaired users, identifying critical sounds such as fire alarms, doorbells, and appliance alerts in real time. By analyzing acoustic patterns, filtering noise, and recognizing distinctive alarm signatures, the system provides timely visual or haptic notifications. This enhances personal safety and independence by ensuring that critical auditory cues are never missed, even in noisy or inaccessible environments. 

Igor Adamenko, Orpaz Ben-Aharon,2025

An extreme image de-warping system based on diffusion models, capable of correcting severe geometric distortions in photos captured with wide-angle lenses, curved surfaces, or unconventional viewpoints. By iteratively refining structure and restoring spatial consistency, the model recovers realistic scene geometry and fine details that traditional de-warping methods struggle to handle. This enables more precise visualization, improved downstream analysis, and higher-quality images in challenging real-world capture conditions. 

Adir Cohen,2025

The project develops a photo-tagging system that leverages graph-based propagation over an image-similarity network. Each photo is represented as a node, and edges encode the degree of visual similarity between image contents. Initial tags derived from confident predictions or a small labeled subset are then propagated across this graph so that visually similar photos reinforce and refine each other’s labels. This enables more accurate, consistent, and rich tagging of extensive photo collections, even when many images individually have weak or ambiguous visual cues. 

Yuri Liziakin.2025

The project develops an active-vision framework for aerial imaging that captures only the most informative image patches at high resolution, thereby maximizing overall captioning quality for VLM models. A coarse, low-resolution scan is first analyzed to predict which regions are most valuable for semantic understanding, after which the system selectively acquires high-resolution views of those areas. This strategy reduces bandwidth and acquisition time while preserving rich scene detail, enabling more accurate and efficient aerial image captioning. 

Matan Chazanovitz 

The project develops a strategic image-acquisition framework for Visual Question Answering, where the system actively selects which parts of a scene to capture or refine based on the specific question being asked. Instead of acquiring full high-resolution imagery, a coarse initial view is analyzed to determine which regions contain the information most relevant to answering the query. The system then directs targeted, high-detail acquisition of those regions, improving answer accuracy while reducing bandwidth, latency, and computational cost. 

Eden Moran

The project develops a system for implicit entity recognition in text, identifying people, organizations, locations, and other entities that are referred to indirectly rather than named explicitly. By combining contextual reasoning, semantic inference, and discourse-level cues, the model infers hidden or implied entities from descriptions, roles, actions, or relationships. This enables deeper text understanding in domains where entities are frequently mentioned obliquely, improving downstream tasks such as information extraction, summarization, and knowledge base construction. 

Omri Sasson

The project develops a framework for generating imperfect student simulations using LLMs, creating virtual learners that exhibit realistic misunderstandings, partial knowledge, inconsistent reasoning, and common conceptual errors. By modeling not only correct responses but also typical mistakes and learning gaps, the system produces richer training data for tutoring platforms, assessment tools, and classroom-support systems. This enables more robust evaluation of educational technologies and provides instructors with realistic scenarios for testing feedback strategies and adaptive teaching methods. 

Yigal Meshulam

The project develops a decentralized multi-robot task-allocation framework operating under a zero-knowledge assumption, in which robots have no prior information about task characteristics or peer capabilities. Instead, each robot independently selects tasks and learns solely by observing its own outcomes. Using a collaborative-filtering-inspired approach, robots share minimal performance signals to infer which task–robot pairings are most effective, gradually converging toward efficient global allocation without central coordination or explicit communication of internal models. This results in scalable, resilient task assignment suitable for environments with uncertainty and dynamism.