Aerial scene understanding
Segmentation, depth, geolocalization and safe-landing estimation from drone and satellite images, trained with as few labels as we can get away with.
15 papersResearch areas
The group started from unsupervised learning in video and graph-based optimization. Most current projects grew out of that work, often with drones as the platform and with partners at NORCE, Google and Bitdefender.
Multi-task hypergraphs
Many networks, each predicting one view of a scene from another, that keep training on unlabelled video by agreeing with each other.
8 papers
Autonomous drones
Small, cheap drones that fly, navigate and land using a single camera.
6 papers
Human–AI interaction and art
Real-time, non-verbal interaction between people and AI, from reading body language to art installations that respond to a smile.
13 papers
Vision and language
Graphs of Events in Space and Time (GEST) as a common representation for video and text, video generation from those graphs, and Romanian language models.
16 papers
Unsupervised learning in space and time
Finding, segmenting and tracking objects in video without labels, using spectral clustering, graph optimization and teacher-student training.
15 papers
Robust and trustworthy machine learning
Anomaly and novelty detection under distribution shift, continual learning, explainability and test-time augmentation.
18 papers
3D reconstruction and novel view synthesis
Combining analytic geometry with neural rendering for novel views, multi-view reconstruction and synthetic 3D data.
4 papers
Earth observation and climate
Hypergraphs over NASA Earth-observation layers that predict one climate variable from the others.
1 paper
Medical imaging
Pulmonary embolism detection and weakly supervised segmentation in CT, vital signs from video, and 3D super-resolution.
5 papersCurrent funding
HRIA: Romanian Hub for Artificial Intelligence
2025–2029Our main source of funding. A national AI hub coordinated by POLITEHNICA Bucharest with other Romanian universities and companies.
Smart Growth, Digitization and Financial Instruments Programme 2021–2027, MySMIS 334906
ELIAS: European Lighthouse of AI for Sustainability
2023–Horizon Europe network of excellence on sustainable AI, under the ELLIS umbrella. The lab takes part together with Bitdefender.
Horizon Europe, grant agreement 101120237
Past projects
SpaceTime Vision: towards unsupervised learning in the 4D world
2019–2022Romania–Norway project with NORCE on unsupervised learning from 3D-plus-time data recorded by fixed sensors and drones.
EEA & Norway Grants EEA-RO-2018-0496 · €1.5M
Robots and Society: Cognitive Systems for Personal Robots and Autonomous Vehicles
2018–2020National project on cognitive systems for personal robots and autonomous vehicles. Marius Leordeanu led the IMAR partner.
UEFISCDI PN-III-P1-1.2-PCCDI2017-0734 · €1.7M
Automatic interpretation of images and video using natural language processing
2015–2019Led together with Traian Rebedea. Most of the early aerial work (road maps, geolocalization, SafeUAV, Ruralscapes) was done in this project.
European Funds POC-A1.2.1D-2015-P39-287 · €1M
Vision in Words: automatic linguistic description of objects, people and their interactions in indoor videos
2018–2020Video-to-language description of indoor scenes.
UEFISCDI TE-2016-2182 · €100K
The Classifier Graph: a recursive multiclass network for deep category recognition in images and video
2016–2018Recursive graphs of classifiers for category recognition.
UEFISCDI ERC-like grant ERC-2016-0007 · €170K
Automatic linguistic descriptions of objects, people and their interactions in indoor videos
2016–2018Experimental demonstrative project on automatic video description.
UEFISCDI PED-2016-1842 · €105K
Automatic video understanding at middle and higher levels of interpretation
2012–2016Exploratory research project on video understanding.
UEFISCDI PCE-2012-4-0581 · €300K