Graduate Projects:

We always look for excellent graduate students or postdocs to join the lab, (including interns from the Computer Science department or the School of Brain and cognitive sciences, the Dkalim, Ashalim, and Eitan programs). If any of the vision sciences catches your imagination and research is your passion, please do contact us to tell us about yourself and learn more about the lab and its research. Research projects for the 2024/2025 academic year are described in this page. Unless stated differently, most projects offer opportunities at all of the M.Sc., Ph.D., and Postdoc levels. For details please contact: Prof. Ohad Ben-Shahar

Algorithmic visual puzzle solving

Computational solving of jigsaw puzzles has been receiving increasing interest in the computational vision community, and our lab has been pushing the abilities of such algorithms to new fronts. But there is still a gap between theoretical/synthetic puzzle solving to scenarios of greater complexity and to real-life applications. Several different research projects are available under this rubric though strong background in computer science, machine learning (in particular deep learning), algorithms, and computational geometry is a plus in all.

  • Real-life algorithmic puzzle solving with applications for computational archeology: One particularly appealing application domain for real-life puzzle-solving challenges is computational cultural heritage, and more specifically, computational archeology. In such cases, very challenging jigsaw puzzle-like problems arise naturally, where fragments are shaped arbitrarily, they are degraded both geometrically and pictorially, some fragments may be missing altogether, while others may be redundant or unnecessary. Can we solve such a computational problem effectively? Involvement in this line research includes collaboration with international teams and true real-life computational archeology challenges that no human was able to solve.
  • Computational theoreical analysis of convex partition puzzles: Among the different puzzle types formulated and addressed in the ICVL are convex partition puzzles, characterized by fragmentation of a visual whole into convex but otherwise arbitrary fragments. The generation process of such puzzles is based on a random set of seed points, that serve as junctions of vertices of the convex fragmentation of their convex hull. An interesting property of such a process is the number of partitions that can be generated for a given seed set, where it is knows that their number, but also their locations, significantly affects it. While a constructive algorithm that can enumerate all convex partitions of a given seed set was already developed, it is current the only known way to count their number, clearly an inefficient way of doing so. In this mostly theoretical project we will try to derive theoretical (upper and lower) bounds for this number. Background and interest in computational geometry and algorithms is an advantage.
  • Puzzle solving for tiny fragments: Automatic solving of jigsaw puzzles, square puzzles in particular , has been studied for many years and obtained impressive results, as long as the fragments are large enough and incorporate enough pictorial information. But experience also show that this capacity deteriorate rapidly when the fragments become smaller and smaller. How far can we push this limit is a matter of research in this project, with the aim of reaching as close as posisble to (the likely unreachable bound of) single pixel fragments. Background in algorithms and deep learning methods likely a plus. Note that there are interesting links between this project and the one on the phase problem in natural image statistics.
  • Getting "real" with square jigsaw puzzles: While automatic solutions for square visual puzzles (such as ours) have obtained impressive results in the literature, they are invariably studied with synthetic fragments cut out perfectly from images stored in the computer memory. But in reality, fragments will be acquired from physical pieces and will suffer from noise, at the very least due to aliasing and misalignment with the pixel grid. What is the implications of "geting real" with these fragments? We will first study this problem by generating proper datasets of such puzzles and testing a plethora of algorithmic solutions designed for the "sterile" data commonly used. Assuming performance is degraded significantly comapred to sterile puzzles, we will then explore new ways to obtain practical and useful results. .

Dolittle AI - Artificial Intelligenc for animal language understanding

Speaking with animals has always been a fundamental human desire. Understanding non-human communication also has great societal and commercial benefits, including communicating with our pets, assessing the ‘mood’ of farm animals, biodiversity protection and conservation, decoding infant communication, and one day even communicating with extra-terrestrial life forms. Existing AI tools, LLMs in particular, are geared toward human language, leading to very poor performance on animal language. Joining forces with interdisciplinary teams from other Israeli universities, and seeking a multimodal approach that combines vision also, in this project we seek to make significant progress on animal language AI. Background in Computer Science, Machine and Deep learning, and Computer vision, are all a plus.

Visual navigation in fish

Navigation is a fundamental behavior deployed across the animal kingdom and is a key capacity for survival. While early life forms utilized random movements, phototaxis, and possibly other elementary sensory-guided movement behaviors, navigation in higher level animals, certainly most vertebrates, typically means the planning and execution of a set of movements in order to realize a predefined path or to find one's way to a destination in some (often familiar) territory. While much visual navigation-related research was done in mammals (rodents in particular), here we will focus on fish, on which there is very limited literature. In this project we will use highly sophisticated robotic platforms and remotely controlled laboratory setups in carefully crafted behavioral experiments, where Flowerhorn Cichlids, an allegedly intelligent and alert teleost fish, will be confronted with progressively more complex visual navigation tasks to reveal its inner spatial representation and navigation "algorithms". Strong background in Brain and Cognitive Sciences is a plus but not a must.

Biological Robots (biobots)

Constructing autonomous physical agents (robots) is a challenging engineering task, especially if they need to act in (possibly hostile) natural environments based on programmable task specifications, while maintaining high level of durability, survivability, and energy independence for prolonged times. The need for autonomy is paramount especially when communication with a remote human operator is highly delayed (think of the Mars rover) or prohibited altogether (think of deep sea missions). In this project we seek to address this challenge with autonomous biobots, namely biological robots whose behavior is fully programmable and controlled via brain stimulation by artificial intelligence and sensory data. Within that general framework, the goal of this project is to develop an aquatic biobot that is based on fish and can exercise desired locomotion in the aquatic environment while performing programmable tasks (as opposed to its natural instinctive behavior).

Computational vision for the study of fish behavior in the wild

Computational tools often play a key component in the study of animal behavior. Often in collaboration with ecologists, zoologists, and physiologists, we study both behavior and neurophysiology of animals in their natural habitat. This is particularly challenging when it comes to marine animals, as the environment is challenging in terms of the methodological and technological tools that should bne employed. In this project we seek to study fish behavior in the wild (in this case, the coral reef in Eilat), by developing fully autonomic robotic tools and algorithmic solutions for the acoustic and visual localization, tracking of fish, we well as visual analysis of the environment, in order to explore and address scientific questions about fish behavior.

Computational vision for aquaculture

Using computer vision for aquaculture holds immense potential to revolutionize how we manage and sustain aquatic farming. This research project seeks to leverage advanced computational vision techniques, including machine learning and image processing, to develop innovative solutions for critical aquaculture challenges. Specifically, we're partnering with the National Aquaculture Center to devise new ways of tracking and analyzing larval breeding. These improvements aim to boost growing system efficiency and yield. The project may require occasional travel to Eilat. Background in computer vision is an advantage but not a must.

Computational visual contour completion

Visual completion is a fundamental perception capacity of human vision and a crucial function for computer vision. Previous computational research in the iCVL has explored the visual contour completion problem in a biologically-inspired and perceptual-consistent fashion, pushing a unique approach that considers the problem not in the image plane but rather in a mathematical space that abstracts the primary visual cortex. Follow up projects include both computational, perceptual, and applied aspects of the problem.

  • A biologically inspired completion approach with curvature in mind: While previous completion models we developed in the lab either ignored findings about curvature tuning of cells in the primary visual cortex, or considered it only implicitly, in this project we seek to develop the theory that considers such curvature tuning explicitly. This should result in an ultimate curve completion theory, that we will also explore and try to validate perceptually. No special background is needed, but previous experience with tools from analysis, differential equations, differential geometry, and numerical analysis, as well as background in visual perception, are a plus.
  • The role of the occluder on amodal completion: Previous perceptual research in the iCVL and worldwide has explored the effect of various visual cues, and in particular the geometric properties of the inducers, on the perceptual task, most often the shape od the completed contours. In this research project we would like to explore how this shape is affected bu the occluder. In addition to its shape, we are interested to understand if the shading pattern on the occluder has a role in the perceptual completion outcome. Strong background in Brain and Cognitive Sciences is a plus but not a must.
  • Shape completion in neglect patients: Visual completion is a fundamental perceptual capacity that allows human observers experience whole and coherent objects from visual fragments. Does this capacity depend on the integrity of spatial vision? In this project we will explore this question by testing neglect patients who suffer from a particular spatial deficit related to their spatial visual field. Key in this project is the formulation of special stimuli in which normal and neglect patients are predicted to make different completions. Strong background in Brain and Cognitive Sciences is an advantage.

Eye-movements exploration

Humans execute rapid eye movements between locations in the visual field well over 100,000 times a day. These movements are an integral part of our visual system, serving to acquire relevant information, track visual targets, and allow us to operate effectively in our world. Tracking eye-movements and measuring how they change in response to different visual stimuli provide a window into human visual perception. Using either of the ICVL’s head mounted, remote, and portable eye trackers, we are able to explore various aspects of these eye movements and utilize them for the understanding of human behavior for various computational, scientific, and even clinical/cognitive applications. Projects involving either computational, behavioral, and cognitive methods are available, and while prior background in eye movements is not necessary, it is a plus.

  • A better scan path segmentation: Eye tracking devices measure raw gaze movements, which are then typically segmented or parsed into fixations and saccades based on de-facto standards set by their manufacturers. These parsed data points collectively form a scan path that serves countless studies and applications. But unfortunately, there is no one single, universally ratified international standard for fixation/saccade algorithms, making the parsing of raw gaze points into scan paths a matter of long-lived debate in the community. While aimed mostly at interns, in this research we will explore the implications of the different common algorithms on the scientific insights obtained from scan paths, seek a critical perspective, and try to develop a better parser.
  • Deep vs. handcrafted eye movement features: The analysis of eye movements, and the representation of either scan paths or raw patterns, typically relies on hand crafted features that are computed from the eye movement time series. In fact, researchers have devised hundreds of such features, either global or local. Not unlike features in other domains, in this project we will explore the utility of eye movement features that are extracted automatically by deep learning models. Can these feature perform better at analyzing eye movement patterns? Can they be interpreted by and for humans to reveal new properties of gaze sequences or scan paths? Background in machine and deep learning can be an advantage.
  • Scale space analysis of raw eye movements: One alternative to scan path segmentation of raw eye movement data is to avoid such parsing altogether and consider the raw data directly. This somewhat naive yet revolutionary possibility can further be considered in a multi-scale approach to highlight events of various temporal scales. In this research we will explore this approach, possibly leading to a new paradign in eye movement representation and analysis.
  • Global spatial modeling of eye movement biases (based on the Polar Saccadic FLow) One aspect of eye movements we study in the iCVL concerns biases in the movement pattern. A recent statistical model we developed, dubbed the Polar Saccadic Flow, provides state-of-the-art modeling of these biases based on the location of saccades launch sites in the visual field. In other words, given the current fixation location, the model provides a probability distribution for the direction and magnitude (size) of the forthcoming saccade. And yet, this prediction is based on a collection of spatially local models, where a single global model is much more desired. This is the goal of this research, which likely to involve tools from optimization and statistics , in addition to experimentation with eye movement acquisition and and analysis.
  • Eye movement analysis for early screening of cognitive and clinical anomalies The eyes are the window to the soul, so says an old proverb. However, more concretely, they are often a window to our body, providing certain early warnings to non-visual physical health problems, from brain tumors, Multiple Sclerosis, diabetes,and high cholesterol, to name a few. Similarly, though not as established, the eyes are a window for cognitive abnormalities, for example cognitive decline due to Alzheimer's disease. In this line of research we seek to explore eye movements for early detection and screening or various cognitive and clinical conditions, from autism, through Dyslexia, to myopia. Background in Brain and Cognitive sciences is an advantage, but not mandatory.

The perceptual hierarchy of scene and object recognition

The ability of humans to recognize and classify both visual scenes and objects is astonishingly rapid and reliable, even when the number of examples experienced from a particular class is very small. Part of this capacity may be assisted by a hierarchical representation of visual classes, reminiscent of artificial datasets like ImageNet, and computational foundations represented as ontologies of the sort our lab has been studying for some time. While ImageNet is based on hierarchy crafted from lexical relationships, a question remains about the perceptual hierarchy that may be innate to the human visual system and how it may have formed during development. In this research project we will explore this question, first using psychophysical and behavioral tools, followed by a computational inquiry that could eventually lead to better recognition algorithms. For this project an interdisciplinary background in both Cognitive/behavioral Sciences and Computational Sciences is a plus, thouhg not a must.

Hyperspectral vision

While color vision allows humans and other primates to see the world in amazing detail, light in the visible spectrum carries more information than just “red, green, and blue”. Advanced “Hyperspectral” cameras allow us to explore the would around us in greater spectral details, revealing the hidden "colors" that lie beyond the capabilities of human vision. Following initial research in the lab that results with new technology for hyperspectral acquisition and reconstruction, several research projects are available about the links between hyperspectral images, color images, and vision.

The phase problem in natural image statistics

The statistics of natural images is a source for great advancements in computer vision, as well as in computational modeling of biological and human vision. In this research project we will explore the implications of certain statistical properties of images for a problem closely related to the Phase Problem in crystallography and investigate applications for image representation, compression, retrieval and search operations.


Last updated: 16 July 2025