- artificial intelligence
- computer vision
- deep learning
Adrien is a Research Scientist at the Xerox Research Centre Europe (XRCE) in the Computer Vision group. His research interests lie in the fields of Computer Vision and Machine Learning, with a focus on automatic video understanding (e.g. behavior recognition, motion analysis, event detection) and object recognition.
- Co-organizer of the first international workshop on Virtual/Augmented Reality for Visual Artificial Intelligence (VARVAI) at ECCV 2016
- One paper accepted at ECCV 2016 on "Sympathy for the Details: Dense Trajectories and Hybrid Classification Architectures for Action Recognition"
- One paper and demo accepted at CVPR 2016 on "Virtual Worlds as Proxy for Multi-Object Tracking Analysis", and new "Virtual KITTI" dataset released
Press coverage: MIT Tech Review (“To Get Truly Smart, AI Might Need to Play More Video Games”), Wired (“Making AI Play Lots of Videogames Could Be Huge (No, Seriously)”), and Forbes (“How Deep Learning Networks Can Use Virtual Worlds To Solve Real World Problems”).
- A paper on "Extending Generic BPM with Computer Vision Capabilities" with Adrian Mos, and Eleonora Vig accepted at RMSOC 2015
- Two papers accepted at BMVC 2015: deep fishing (extracting gradient features from ConvNets, with Albert Gordo and Florent Perronnin), and ODAMOT (Online Domain Adaptation for Multi-Object Tracking, with Eleonora Vig), which was accepted as oral (7% acceptance ratio).
- Outstanding reviewer award at CVPR 2015.
Video Analytics in a Virtual World
As the gap between real and virtual worlds continues to close, more and more research in computer vision turns to photorealistic imagery to help tackle low- and high-level visual tasks. Eleonora Vig and myself were awarded a Xerox internal grant (XTIN) for ambitious exploratory research combining Computer Vision and Computer Graphics. In particular, we are exploring the use of Virtual Worlds for high-level Computer Vision tasks.
Visual Understanding of Processes
I am currently leading an ambitious project aiming to combine Computer Vision (scene understanding, multi-object tracking, action recognition, ...) with Business Process Modeling (BPM). The main goal is to increase business agility via the automatic integration of visual AI components in BPM. This project is a collaboration with Adrian Mos, Naila Murray, Eleonora Vig, Diane Larlus, Albert Gordo, and Jon Almazan.
Learning Representations for Video Analytics
I am also involved in a project aiming to learn representations for a variety of multimedia analytics tasks (speech recognition and various computer vision problems). In particular, I study the problem of learning spatio-temporal representations (for instance via deep learning). Together with Antonio M. López and Eleonora Vig, we are co-supervising a PhD student (Cesar De Souza) on this topic. In another direction related to deep learning, I also did some work on accelerating SGD training by online learning to sample together with Guillaume Bouchard, Theo Trouillon, and Julien Perez.
You can find most of my publications through Google Scholar
- Code for my IJCV'14 paper on Activity representation with motion hierarchies:
- Jupyter notebook to compute Actom Sequence Models (ASM) for action recognition in videos:
camocomp, a Python package for camera motion compensation (used in our motion hierarchy paper, and related to improved dense trajectories):
ekovof, a C/Cython/Python package for Efficient Kernels on Vectors of Floats (useful tools for kernel methods):
daco, a Python package to compute the Difference between Auto-Correlation Operators (DACO) and other distance functions between time series of sparse vectors (used in our kernel on time series paper):
- Python wrapper around the excellent Global Alignment kernel code of Marco Cuturi
- some older code snippets you might find useful: projection on the simplex, a multi-dimensional extension of numpy's digitize, and more.
Adrien graduated from ENSIMAG and obtained an MSc in Artificial Intelligence from Université Joseph Fourier in 2008. He, then, worked as a doctoral candidate at the MSR-INRIA joint center in Paris, and in the LEAR team at INRIA Grenoble under the supervision of Zaid Harchaoui and Cordelia Schmid. He got his PhD from Université de Grenoble in 2012 and joined XRCE in 2013.
Old LEAR webpage: http://lear.inrialpes.fr/people/gaidon/