CVPR 2014 Tutorial on Large-Scale Visual Recognition

Saturday, June 28th - Full day - Grand Ballroom 2

Speakers:

Tutorial goals

This tutorial addresses Large-Scale Visual Recognition (LSVR), the problem of understanding visual content (e.g. photos or videos) on a large-scale. This is a topic which has received much attention in the computer vision community in the last few years: as larger datasets have become available [TFF08, DDS09], handling millions of images and thousands of label classes has become the norm rather than the exception [DBL10, WBU10, LRM12, DCM12, JPD12]. Since LSVR is a vast topic, we will mainly focus on two tasks: image retrieval and image classification.

The goals of this tutorial are three-fold:

• Provide the audience with the "tools" to process such large datasets.
• Show the convergence between large-scale retrieval and large-scale classification, two problems which have been traditionally addressed separately.
• Show that LSVR does not necessarily require massive computational resources (although such resources can help, of course...)

The tutorial is complemented with free publicly available software:

• VLFeat: http://www.vlfeat.org/
• INRIA's Fisher vector implementation: http://lear.inrialpes.fr/src/inria_fisher
• VGG's encoding methods evaluation toolkit: http://www.robots.ox.ac.uk/~vgg/software/enceval_toolkit/
• Yael library for exact matching: http://gforge.inria.fr/projects/yael
  • PQ-codes toy Matlab implementation: http://people.rennes.inria.fr/Herve.Jegou/projects/ann.html
• J-SGD for large-scale learning: http://lear.inrialpes.fr/src/jsgd/

Schedule

The tutorial will consist of short talks (1h or less) each one covering a specific topic, and each one given by a recognized expert in his field.

morning:

• 8:30am - 8:40am: Introduction (Zaid Harchaoui)
• 8:40am - 9:25am: Part I: Efficient matching (Herve Jegou)
• 9:25am - 10:15am: Part II: Geometry for large-scale retrieval (Ondrej Chum)
• 10:15am - 10:45am: COFFEE BREAK
• 10:45am - 11:50am: Part III: Large-scale machine learning (Zaid Harchaoui)

afternoon:

• 1:30pm - 2:30pm: Part IV: Large-scale visual recognition with deep learning (Marc'Aurelio Ranzato)
• 2:30pm - 3:25pm: Part V: Input embeddings, from shallow to deep (Andrea Vedaldi)
• 3:25pm - 3:55pm: COFFEE BREAK
• 3:55pm - 4:55pm: Part VI: Output embedding for large-scale visual recognition (Florent Perronnin)

Here is a list of references.

Sponsors and financial support

The tutorial is supported by the MSR-INRIA Joint Centre, the "Gargantua" project (CNRS-Mastodons), the "Khronos" project (Labex Persyval-Lab, ANR-11-LABX-0025), and the Fire-ID project (ANR-12-CORD-016).