Computer Vision

Machines with the ability to ‘feel’ currently in development as we enter next frontier of AI

Is that solar panel pointing in the right direction?

A new technique uses a single image to forecast solar panel energy production and maximize output. The method estimates the amount of energy that will be produced based on the angle of the sun, shadows, reflections, and weather patterns, allowing for more accurate placement and optimization of solar panels in urban areas.

NTU Singapore scientists invent AI-powered biochip that detects genetic markers in 20 minutes

A team of scientists from NTU Singapore has developed a new biochip that, when paired with Artificial Intelligence (AI), can detect quickly and accurately extremely small amounts of microRNAs. The device can cut detection time from hours to 20 minutes.

HEAPGrasp: A faster, smarter way for robots to handle tricky objects using only RGB camera

HeapGrasp uses RGB images to analyze object silhouettes and estimate its 3D shape, reducing the need for depth information. The approach achieves high accuracy while minimizing camera movement and execution time.

KTU researchers develop a model that improves machine understanding of the real world

A new model combines multiple ways of analysing 3D data, integrating local and global perspectives to interpret complex environments more reliably. The system improves detection of small or partially visible objects in real-world situations, enhancing safety in autonomous systems.

Improving AI models’ ability to explain their predictions

Researchers at MIT developed a new method that coaxes AI models to achieve better accuracy and clearer explanations in safety-critical applications. The approach extracts concepts the model has learned while training for a specific task and forces it to use those, producing better explanations than standard concept bottleneck models.

Small models, big insights into vision

Researchers used machine learning techniques to compress a large model of the visual cortex, creating smaller versions that predict neural responses with high accuracy. The compact models revealed specific computational patterns in how neurons detect important features, offering insights into how visual information is processed.

Philadelphia communities help AI computer vision get better at spotting gentrification

Drexel researchers create machine learning program that integrates qualitative and quantitative data to identify gentrification in Philadelphia neighborhoods. The program, trained with data from thousands of images and focus groups, accurately identifies new-build gentrification with 84% accuracy.

InstaDrive: Street view generation based on the unified instance segmentation input of vehicles and map elements

InstaDrive generates precise editing of vehicles and map elements, enabling efficient labeled data generation. It outperforms baselines in FID and mAP, preserving accurate map structures and maintaining multi-view consistency.

Deep-learning model predicts how fruit flies form, cell by cell

A team of MIT engineers developed a deep-learning model that predicts how individual cells will fold, divide, and rearrange during a fruit fly's earliest stage of growth. The model achieved 90% accuracy in predicting the movement of 5,000 cells over the first hour of development.

Can AI read humans’ minds? A new model shows it’s shockingly good at it

A breakthrough AI system called OmniPredict can predict human pedestrian behaviors with unprecedented accuracy, revolutionizing self-driving cars and urban mobility. The model combines visual cues with contextual information to anticipate pedestrians' next moves, reducing the risk of accidents and improving traffic safety.

AI-powered vision gives meaning to wildfire chaos

A UBC Okanagan team harnesses computer modeling to study wildfire movement, finding that fires often behave randomly due to factors like fuel type, wind, and terrain. This randomness can lead to significant variations in fire spread, highlighting the need for more probabilistic models.

Purdue innovation to be evaluated in international study for earlier identification of preeclampsia risk

Researchers at Purdue University are testing a computer-vision method to analyze smartphone photos of pregnant women's eyes to predict preeclampsia risk. The two-year study aims to reduce maternal mortality in Africa and could potentially save thousands of lives.

AI at the Eyelid: Glasses that track health through your blinks

Researchers developed AI-powered BlinkWise glasses that track blinking patterns to assess fatigue, mental workload, and eye-related health issues. The device uses radio signals to detect minute eyelid movements with unprecedented detail, preserving privacy and using minimal power.

‘More than just an image’: Purdue tech extracts hyperspectral info from conventional photos

Researchers at Purdue University have developed an algorithm that recovers detailed spectral information from photographs taken by conventional cameras. The method uses computer vision, color science, and optical spectroscopy to achieve high spectral resolution comparable to scientific spectrometers.

AI turns printer into a partner in tissue engineering

Researchers at UMC Utrecht developed a new AI-powered printer called GRACE that can print implantable tissues with improved cell survival and functionality. The printer uses computer vision and laser-based imaging to design and print complex structures, including blood vessels and cartilage layers.

Shaky cameras can make for sharper shots, new research shows

Researchers at Brown University developed an image processing technique that harnesses camera motion to increase resolution, producing super-resolution images with details sharper than the original pixel array allows. The technique has potential applications in archival photography and photography from moving aircraft.

An efficient and memory-friendly unsupervised industrial anomaly detection model

A research team developed an innovative unsupervised model for industrial anomaly detection using paired well-lit and low-light images. The model leverages feature maps, Low-pass Feature Enhancement, and Illumination-aware Feature Enhancement to detect anomalies while remaining lightweight and memory-efficient.

Pedestrians now walk faster and linger less, researchers find

A new study reveals that pedestrians are now walking faster and spending less time in public spaces. Researchers analyzed 40 years of video footage to find a 14% decline in people lingering in these areas.

AI vision, reinvented: The power of synthetic data

Researchers developed CoSyn, a new approach to train open-source models using AI-generated scientific figures and charts. The resulting dataset, CoSyn-400K, includes over 400,000 synthetic images and 2.7 million sets of corresponding instructions. CoSyn-trained models match or outperform proprietary peers in various benchmark tests.

New tool gives anyone the ability to train a robot

MIT engineers developed a versatile demonstration interface that allows users to teach robots new skills in three intuitive ways: remote control, physical manipulation, or demonstration. This innovation expands the type of users and 'teachers' who interact with robots, enabling robots to learn a wider set of skills.

New attack can make AI ‘see’ whatever you want

Researchers have demonstrated a new technique, RisingAttacK, to manipulate all widely used AI computer vision systems, allowing them to control what the AI 'sees'. The attack is effective at influencing the AI's ability to detect top targets, such as cars, pedestrians, or stop signs.

Pervasive surveillance of people is being used to access, monetize, coerce, and control

A new study reveals a five-fold increase in computer vision papers linked to surveillance patents, highlighting the rise of obfuscating language that normalises surveillance. The top institutions producing surveillance are Microsoft, Carnegie Mellon University, and MIT.

New all-silicon computer vision hardware by UMass researchers advances in-sensor visual processing technology

Researchers at UMass Amherst created integrated arrays of gate-tunable silicon photodetectors that can capture dynamic visual information and classify static images with high accuracy. The technology has the potential to reduce latency in computer vision tasks, enabling applications like self-driving vehicles and bioimaging.

Animation technique simulates the motion of squishy objects

Researchers at MIT developed a simulation method that allows for accurate and stable simulations of elastic materials, enabling the creation of realistic bouncy characters in movies and video games. The approach preserves physical properties and avoids instability, making it a promising tool for engineers to design flexible products.

Imaging technique removes the effect of water in underwater scenes

Researchers have developed an image-analysis tool called SeaSplat that cuts through the ocean's optical effects and generates images of underwater environments with accurate colors. The team paired SeaSplat with a computational model to convert images into three-dimensional underwater worlds, allowing for virtual exploration.

Smarter skies: A new AI model turns street cameras into rainfall sensors

Researchers developed an innovative deep-learning-based framework that uses common surveillance cameras to estimate rainfall in real time. The approach achieved high predictive accuracy across various environmental conditions and lighting scenarios, outperforming traditional methods while maintaining low computational costs.

New AI model dramatically improves subgraph matching accuracy by eliminating noise

A new deep learning model, ENDNet, significantly enhances subgraph matching accuracy by identifying and neutralizing extra nodes that interfere with the matching process. This improves performance in pattern recognition tasks across various fields, including drug discovery and natural language processing.

Making AI models more trustworthy for high-stakes settings

Researchers at MIT developed a technique to improve the reliability of conformal classification, which can produce impractably large prediction sets. By combining test-time augmentation with conformal prediction, they reduced prediction set sizes by up to 30 percent while maintaining probability guarantees.

Helping computers perceive and interact with the visual world

Schmid's contributions have helped computers recognize complex objects, understand video analysis, and process realistic settings. Her leadership has built active research communities, mentoring and supervising peers across the field of computer vision.

UF professor develops AI tool to better assess Parkinson’s disease, other movement disorders

A University of Florida researcher has developed a groundbreaking AI tool called VisionMD that analyzes videos of patients with Parkinson's disease and other movement disorders. The tool provides valuable information about how the disease is progressing and responding to medications, improving patient care and advancing clinical research.

Transforming doors into gateways to the virtual world: the future of mixed reality!

A collaborative research team has developed a novel mixed reality (MR) technology that uses real-world doors as natural transition points. The system allows users to select a door within their MR interface and seamlessly transition into a virtual space, creating an unprecedented sense of immersion.

New 3D technology paves way for next-generation eye-tracking

Researchers at the University of Arizona have developed a new 3D imaging technique, deflectometry, paired with advanced computation to improve eye-tracking accuracy. The method can capture gaze direction information from more than 40,000 surface points, theoretically millions, increasing accuracy by a factor of over 3,000 compared to c...

Beyond ambiguous reflections: Bridging optical 3D metrology and computer vision

Researchers develop a new approach combining Phase Measuring Deflectometry and Shape from Polarization to accurately image specular surfaces without prior knowledge or assumptions. The single-shot method enables motion-robust measurements, pushing the limits for next-generation 3D sensors.

New AI tool generates high-quality images faster than state-of-the-art approaches

Researchers have developed a hybrid image-generation tool called HART that combines the strengths of autoregressive and diffusion models. It achieves high reconstruction quality with significantly reduced computational resources, enabling local execution on laptops or smartphones.

2025 IEEE 2nd International Conference on Deep Learning and Computer Vision (IEEE DLCV 2025)

The conference aims to bridge theoretical advancements with practical applications in AI and visual computing. Researchers can submit original research papers and attend keynote sessions, offering opportunities to network with pioneers in intelligent technologies.

From handicap to asset: AI approach leverages optics phenomenon to produce better images

Scientists developed a method that harnesses chromatic aberration to produce high-quality images using a single exposure. The AI approach uses generative models to retrieve phase information from limited data input.

Two Michigan Engineering researchers named 2025 Sloan Research Fellows

Thatchaphol Saranurak and Andrew Owens have been awarded Sloan Research Fellowships for their innovative work on graph networks and machine perception systems. Their research aims to create more efficient algorithms for computing dynamic systems, such as social networks and traffic patterns.

Engineers enable a drone to determine its position in the dark and indoors

MIT researchers have introduced a new system called MiFly that enables drones to self-localize in indoor, dark, and low-visibility environments. The system uses radio frequency waves reflected by a single tag placed in the environment, allowing the drone to estimate its trajectory with high accuracy.

How neighborhood perception affects housing rents: A novel analytical approach

A new method developed by Osaka Metropolitan University accurately predicts housing prices in Osaka City, with neighborhood perception being a key factor. The approach achieves nearly 75% accuracy by combining existing property data with machine-learning-processed street view images.

UCF helps develop AI tool that may assist underserved hospitals

The open-source AI model analyzes medical images, generates detailed reports, and answers clinical questions to streamline diagnostics and improve accuracy. BiomedGPT aims to democratize healthcare and reduce disparities amongst patients by providing easily accessible data to bolster underserved hospitals.

A review of camouflaged object detection research and the promise of deep learning

A comprehensive review of camouflaged object detection research highlights the potential of deep learning in recognizing objects in complex scenarios. The review analyzes traditional and deep learning approaches, emphasizing practical contributions and theoretical frameworks.

Smart food drying techniques with AI enhance product quality and efficiency

Researchers develop precision techniques using optical sensors and AI to facilitate efficient and accurate food drying. The study discusses three emerging smart drying techniques, providing practical information for the food industry.

Placenta assessment tool aims to improve neonatal, maternal care

A new tool developed by Penn State researchers uses computer vision and artificial intelligence to analyze placenta images, detecting abnormalities and risks such as neonatal sepsis. The PlacentaCLIP+ model has the potential to transform neonatal and maternal care in low- and high-resource settings.

The best AI strategy to recognize multiple objects in one image

Researchers from Bar-Ilan University discover that classifying objects together through Multi-Label Classification can yield better results than detecting individual objects. This new method allows networks to learn correlations between object combinations, making them more recognizable in real-life applications such as autonomous vehi...

UTIA team wins grant to advance AI education and career preparation

The University of Tennessee Institute of Agriculture has won a four-year grant to create hands-on curriculum about AI-related technologies for future farmers and leaders. Selected students will test the curriculum in drones, robotics, and other smart agriculture technologies, gaining skills in coding, drone-work, and robotics.

A new way to create realistic 3D shapes using generative AI

Researchers develop a simple fix to an existing technique, enabling the generation of sharp, high-quality 3D shapes that rival top model-generated 2D images. The new approach improves upon previous methods by avoiding costly retraining and complex postprocessing.

Researchers develop markerless motion capture system to push biomechanics “into the wild”

Researchers at CAMERA have developed an open-source markerless motion capture system using computer vision and deep learning methods. The system estimates joint positions from regular 2D image data, providing unobtrusive analysis of body movements.

Reality check: making indoor smartphone-based augmented reality work

A study by Osaka University researchers found that visual landmarks can be difficult to find in certain environments, leading to motion sickness. They propose using radio-frequency localization, such as ultra-wideband sensing, to overcome these challenges and improve indoor augmented reality applications.

Next step in light microscopy image improvement

A new computational model called Multi-Stage Residual-BCR Net (m-rBCR) uses a unique frequency representation to solve deconvolution tasks with fewer parameters and faster processing times. The model demonstrates high performance on various microscopy datasets, outperforming traditional methods.

Could crowdsourcing hold the key to early wildfire detection?

A new crowdsourcing system, FireLoc, uses a network of low-cost mobile phones to detect wildfires minutes—even seconds—after they ignite. The system prioritizes privacy and accurately maps wilderness fires to within 180 feet of their origin.

KAIST proposes AI training method that will drastically shorten time for complex quantum mechanical calculations​

Researchers developed a novel AI approach to predict atomic-level chemical bonding information in 3D space, bypassing traditional supercomputer simulations. This methodology accelerates calculations by learning chemical bonding information using neural network algorithms from computer vision.

With the help of AI, UC Berkeley researchers confirm Hollywood is getting more diverse

Researchers used facial recognition technology to track actor screen time in over 2,300 films, confirming a shift towards greater diversity. The study found that individual film casts are becoming more diverse, with non-leading roles exhibiting more variety than leading ones.

Real-time descriptions of surroundings for people who are blind

WorldScribe, a new software, uses generative AI to provide real-time text and audio descriptions of surroundings for people who are blind or have low vision. The tool can adjust the level of detail based on user commands or camera frame time.

Helping robots zero in on the objects that matter

A new method called Clio allows robots to make task-relevant decisions by identifying the parts of a scene that matter. In real experiments, Clio successfully mapped scenes at different levels of granularity based on natural-language prompts and enabled robots to grasp objects of interest.

Rice research could make weird AI images a thing of the past

Rice University researchers developed ElasticDiffusion, a method that separates local and global signals to create non-square aspect ratio images without visual imperfections. The new approach can improve consistency and realism in AI-generated images, but still requires significant computational power.

Bilateral reference framework for high-resolution dichotomous image segmentation

Researchers at Tsinghua University Press have developed BiRefNet, a bilateral reference framework that captures tiny-pixel features and achieves highly accurate high-resolution salient object detection and concealed object detection. The framework has numerous practical applications in various fields.

Beetle that pushes dung with the help of 100 billion stars unlocks the key to better navigation systems in drones and satellites

Researchers at the University of South Australia have developed an AI sensor that can accurately measure the orientation of the Milky Way in low light, using a technique inspired by the dung beetle. This system could improve navigation for drones and satellites in difficult lighting conditions.

Researchers develop stress-free method to weigh mice using computer vision

Researchers at Jackson Laboratory have developed a non-intrusive method to accurately and continuously measure mouse body mass using computer vision. This approach reduces stress associated with traditional weighing techniques, improving data accuracy and reproducibility.

Potential and prospects of segment anything model: a survey

The Segment Anything Model has achieved significant breakthroughs in image segmentation, leveraging its data engine methodology and vast datasets. Researchers have proposed improvements and applications for the model, showcasing its versatility across various tasks and domains.