Alessandro's picture

Alessandro Bergamo

Senior Research Scientist
Amazon, Seattle, WA
Email: bergamoale (at) gmail (dot) com
Curriculum Vitae, Google scholar, Linkedin

Welcome · Benvenuto · خوش آمدید · 欢迎 · Bienvenida · ברוך הבא · ترحيب · Accueil · Chào mừng · 歓迎 · добро пожаловать · Willkommen · स्वागत · 환영 · Boas-vindas · kuwakaribisha

My name is Alessandro. I received my PhD from Dartmouth College, under the supervision of Prof. Lorenzo Torresani.
I was part of the Vision Learning Group.
I am now working as Research Scientist at Amazon at their Seattle headquaters.

Research

I am broadly interested in both Machine Learning and Computer Vision, and especially in their interesection. In the past few years I have been working on many aspects of image understanding, recognition, and detection. Part of my PhD research was devoted on designing learning methods to build compact but powerful image representations, subsequentially used for efficient image recognition and image search on large-scale photos collections. I have worked on ad-hoc methods designed for the recognition of landmarks from single user photos. I also worked on domain adaptation techniques, to adapt previsouly-trained classification models to new environments.

Publications

Self-taught object localization with deep networks
Alessandro Bergamo, L. Bazzani, D. Anguelov, L. Torresani
Winter Conference on Applications of Computer Vision (WACV), 2016
[PDF, code]
Classemes and Other Classifier-based Features for Efficient Object Categorization
Alessandro Bergamo, Lorenzo Torresani
Transactions on Pattern Analysis and Machine Intelligence (TPAMI), March 2014
[PDF, supplementary material, code]
AutoCaption: Automatic Caption Generation for Personal Photos
Krishnan Ramnath, Simon Baker, Lucy Vanderwende, Motaz El-Saban, Sudipta Sinha, Anitha Kannan, Noran Hassan, Michel Galley, Yi Yang, Deva Ramanan, Alessandro Bergamo, Lorenzo Torresani
Winter Conference on Applications of Computer Vision (WACV) 2014
[PDF, webpage]
Leveraging Structure from Motion to Learn Discriminative Codebooks for Scalable Landmark Classification
Alessandro Bergamo, Sudipta N. Sinha, Lorenzo Torresani
Computer Vision and Pattern Recognition (CVPR) 2013
[PDF, webpage]
Meta-Class Features for Large-Scale Object Categorization on a Budget
Alessandro Bergamo, Lorenzo Torresani
Computer Vision and Pattern Recognition (CVPR) 2012
[PDF, webpage, code]
PiCoDes: Learning a Compact Code for Novel-Category Recognition
Alessandro Bergamo, Lorenzo Torresani, Andrew Fitzgibbon
Neural Information Processing Systems (NIPS) 2011
[PDF, webpage, code]
Exploiting weakly-labeled Web images to improve object classification: a domain adaptation approach
Alessandro Bergamo, Lorenzo Torresani
Neural Information Processing Systems (NIPS) 2010
[PDF, webpage]
Learning Image Representations for Efficient Recognition of Novel Classes
Alessandro Bergamo, Lorenzo Torresani.
The learning workshop 2011, oral presentation

Code

VLG_extractor

This software extracts the image descriptors described in:

LIBLINEAR_bitmap

This is an improvement of the LIBLINEAR [Fan et al., JMLR 2008] software to efficiently support non-sparse large-scale binary data. Many features have been added, such as multi-core supports via PThreads, several 1-vs-all classification models, down-sampling of the training data, polynomial kernel, b-bit Minwise Hashing as the compression method.
It has been tested on a database with 1.2M of training examples and 100K dimensions on low-budget computers. Written in C/C++ using GCC and PThreads. [code]

Others

Experience

I spent two fantastic Summers doing research for Microsoft Research and Google.
At Microsoft (Jun-Sept 2012) I worked in the Interactive Visual Media Group (Redmond, WA), under the supervision of Sudipta Sinha. I followed a project regarding the automatic recognition of famous landmarks from single photos.

At Google (Jun-Sept 2013), I worked with the Visual Search Group (Mountain View, CA) under the supervision of Dragomir Anguelov. I worked on a project regarding the design of a weakly-supervised visual object detection system, that can be trained with partial annotations.