Building Classification Cascades for
Visual Identification from One Example

Andras Ferencz  UC Berkeley & Mobileye Vision Technologies
Erik G. Learned-Miller  UC Berkeley & UMass Amherst
Jitendra Malik   UC Berkeley
Vidit Jain   UMass Amherst


In this project, we attempt to solve the problem of object identification (OID), which is specialized recognition where the category is known (for example cars or faces) and the algorithm recognizes an object's exact identity (such as Bob's BMW).  For example, we might be given images of cars like these:
A Car Identification Example

and be asked to find which of the 4 cars on right are the same as either of the 2 on the left. Two special challenges characterize OID:
  1. Inter-class variation is often small (many cars look alike) and may be dwarfed by illumination or pose changes.
  2. There may be many classes but few or just one positive ``training'' examples per class.
Due to (1), a solution must locate possibly subtle object-specific salient features (a door handle) while avoiding distracting ones (a specular highlights). However, (2) rules out direct techniques of feature selection.  We describe an on-line algorithm that takes one query image from a known category and builds an efficient ``same'' vs. ``different'' classification cascade by predicting the most discriminative feature set for that object.  Our method not only estimates the saliency and scoring function for each candidate feature, but also models the dependency between features, building an ordered feature sequence unique to a specific query image, maximizing cumulative information content.  Learned stopping thresholds make the classifier very efficient.  To make this possible, category-specific characteristics are learned automatically in an off-line training procedure from labeled image pairs of the category, without prior knowledge about the category.  In functional notation, our algorithm has 3 steps:
  1. (Off-line) trainer Tid: category training images --> identifier generator Hid
  2. (On-line) identifier generator Hid: object image --> classifier Cid
  3. (On-line) classifier Cid: test image --> { same, different }
Where step 1 learns category specific characteristics, while step 2 creates an object specific classifier.


Andras Ferencz, Erik Learned-Miller,  Jitendra Malik.   Learning to Locate Informative Features for Visual Identification.
Draft: submitted to IJCV Special Issue: Learning for Vision, 2007. PDF  

Vidit Jain, Andras Ferencz, Erik Learned-Miller.    Discriminative Training of Hyper-feature Models for Object Identification.
In Proceedings of British Machine Vision Conference, 2006. PDF   [Project Page @ UMass Amherst]

Andras Ferencz, Erik Learned-Miller,  Jitendra Malik.  Building a Classification Cascade for Visual Identification from One Example.
ICCV 2005.  PDF

Andras Ferencz, Erik Learned-Miller,  Jitendra Malik.  Learning Hyper-Features for Visual Identification.
Neural Information Processing Systems, 2004.  PDF


Dissertation Talk. March 2005. PDF  
NIPS*05 Interclass Transfer Workshop. [PDF]


Car and face classification examples (with the matched patches marked). 


Our car data set and an implementation of our system is available for download here


The learning problem that must be solved in visual identification (steps 1 & 2 above) is best demonstrated with an example:

Martian mugshots

If we knew nothing about Martians, finding Bob after having seen only a single image of him before, would be impossible.  However, with a training set of other Martians, even if Bob is not in this set, the problem becomes tractable.

Last Updated 4/27/05 by ferencz