Pasadena, California, United States
7K followers 500+ connections

Join to view profile

About

Leading science for Agentic AI at AWS. Previously led the development of AI Applications…

Activity

Join now to see all activity

Experience & Education

  • Amazon Web Services (AWS)

View Stefano’s full experience

See their title, tenure and more.

or

By clicking Continue to join or sign in, you agree to LinkedIn’s User Agreement, Privacy Policy, and Cookie Policy.

Publications

  • Detachable Object Detection

    A representation of models as detachable surfaces, manifest from occlusion regions.

    See publication

  • On the set of images modulo viewpoint and illumination changes

    Characterizes the maximal invariant of images under the action of viewpoint and illumination changes. This is the invariant that is functioning of all other invariants. It is the quotient of an infinite dimensional space (images, even if of infinite resolution) modulo an infinite-dimensional group (the product of the closure of the set of epipolar diffeomorphisms and range homeomorphisms of the plane), which yield a finite topological object, called Attributed Reeb Tree, that contains all and…

    Characterizes the maximal invariant of images under the action of viewpoint and illumination changes. This is the invariant that is functioning of all other invariants. It is the quotient of an infinite dimensional space (images, even if of infinite resolution) modulo an infinite-dimensional group (the product of the closure of the set of epipolar diffeomorphisms and range homeomorphisms of the plane), which yield a finite topological object, called Attributed Reeb Tree, that contains all and only the information in an image that is unaffected by changes of viewpoint and illumination. It is a discrete, "symbolic" entity that is equivalent to the image, even if it had infinite resolution, if one wishes to remove nuisance variability due to viewpoint and illumination exactly.

    See publication

  • Quickshift and kernel methods for mode seeking

    Quickshift is an unsupervised hierarchical clustering method where data points move towards cluster centers following paths restricted to go through other data points. It was later re-discovered by others http://sites.psu.edu/mcnl/files/2017/03/9-2dhti48.pdf

    See publication

  • KALMANSAC

    Developed first random sample consensus filtering algorithm.

    See publication

  • A semi-direct approach to structure from motion

    First semi-direct method for Structure From Motion (SFM)

    See publication

  • Optimal Structure from Motion: Local Ambiguities and Global Estimates

    The first provably optimal algorithm for SFM (Structural from Motion, AKA Visual SLAM -- simultaneous localization and mapping), Best Paper at CVPR 1998. Introduced CBS Plots, that allow visualizing all local extrema (maxima, minimal, saddles) of the highly non-convex, high-dimensional optimization residual of structure from motion.

    See publication

  • An Invitation to 3D Vision

    -

    A reference textbook for 3D Vision

    See publication

  • Dynamic Textures

    -

    Introduced the first model for dynamic textures

    See publication

  • Real-time 3D motion and structure of point features: a front-end system for vision-based control and interaction

    -

    The first demonstration of real-time SFM (structure from motion, AKA Visual SLAM -- simultaneous localization and mapping). Development started in 1998 and first demonstrated in public at CVPR 2000.

    See publication

  • Virtual Object Insertion

    -

    First demonstration of augmented reality in commodity hardware, ICCV 2001.

    See publication
Join now to see all publications

Patents

Projects

  • Visual Inertial SLAM, VIO, Structure From Motion

    Developed first auto-calibration causal filtering method for VIO (visual inertial odometry), later turned into CORVIS, a filtering-based visual inertial mapping. Corvis then evolved into XIVO https://github.com/ucla-vision/xivo and was incorporated into several commercial systems for visual inertial odometry and SLAM (simultaneous localization and mapping)

    See project

  • Information Theory of Deep Learning

    -

    We have been developing an information-theoretic framework to explain known behavior of deep networks (generalization despite overparametrization, catastrophic forgetting, relation to PAC Bayes bounds, invariance/equivriance of representation)

    See project

  • Visual Inertial Semantic mapping

    -

    The first visual inertial semantic mapping real-time demonstration (CVPR 2017). CORVIS and XIVO integrated inertials into the semantic representations by exploiting global orientation and scale to canonize visual features and increase discriminative power of local descriptors. This project showed that meaningful integration of semantics with inertials allows the system to distinguish real cars from toy cars, as well as to maintain persistent memory of objects that are temporarily occluded, and…

    The first visual inertial semantic mapping real-time demonstration (CVPR 2017). CORVIS and XIVO integrated inertials into the semantic representations by exploiting global orientation and scale to canonize visual features and increase discriminative power of local descriptors. This project showed that meaningful integration of semantics with inertials allows the system to distinguish real cars from toy cars, as well as to maintain persistent memory of objects that are temporarily occluded, and predict when they will become disoccluded. See paper at http://openaccess.thecvf.com/content_cvpr_2017/papers/Dong_Visual-Inertial-Semantic_Scene_Representation_CVPR_2017_paper.pdf , see also https://www.youtube.com/watch?v=TZTriqQm6nU and https://www.youtube.com/watch?v=Rt2jdurowfE

    See project

  • Visual exploration

    -

    An autonomous drone looking for a known object in an unknown environment

    See project

  • Robust Visual Tracking

    -

    Multiscale Template Descriptor for robust visual tracking

    See project

  • Real time object recognition on a hand-held device

    -

    A system to learn and detect objects and object categories all running in real-time on an iPhone 3s

    See project

  • DARPA Grand Challenge

    -

    See project

  • Augmented Reality: First demonstration on commodity hardware

    -

    First demonstration of real-time Structure From Motion (SFM) at CVPR 2000; first demonstration of real-time virtual insertion in live video with a hand-held camera at ICCV 2001.

Honors & Awards

  • ACM Fellow

    Association for Computing Machinery

    For contributions to the foundations of visual geometry and representation learning. Class of 2022.

  • IEEE Fellow

    IEEE

  • Marr Prize

    IEEE

Languages

  • Italian

    -

More activity by Stefano

View Stefano’s full profile

  • See who you know in common
  • Get introduced
  • Contact Stefano directly
Join to view full profile

Other similar profiles

Explore top content on LinkedIn

Find curated posts and insights for relevant topics all in one place.

View top content

Add new skills with these courses

See all courses