Rearview mirror detects familiar faces walking into the camera field of view. Basic idea is to match a query image into a database (representation?). Tough to accomplish because we need to deal with pose variance, and non-rigid expressions.
Pipeline:
- Detect (and align?)
- Build feature vector (spatial pyramid, Fisher vectors, ...) from affine-invariant (engineered) descriptors.
- Apply some dimensionality reduction scheme (possibly learned from training data---metric/similarity learning).
- Match to database image (k-NN?).
- Update database? (Probably later, offline)
Face recognition:
* Taigman, CVPR 2014
* Guillaumin et al. `Is that you? Metric learning approaches for face
identification', ICCV 2009
* Lue et al., `Neighborhood repulsed metric learning for kinship
verification', CVPR 2012
Metric learning:
* Davis, ICML 2007
* Weinberger, NIPS 2006
Back propagation:
* Rumelhart, Nature 1986
* LeCun, IEEE 1998
mlearn implements a few metric learning algorithms.
utils provides some interesting methods for playing around with clustering
and nearest neighbors. Included for computing nearest neighbors (nearest or
fixed-radius) are
- linear :
Linear_nn - cell :
Cell_nn - kd-tree :
Kd_tree - FLANN : wrapper around FLANN for approximate nearest neighbors.
To build:
$ mkdir build
$ cd build
$ cmake ..
$ make
Can we use something like Cao (2013) to match into different regions of a graph for more efficient look up?
For offline labeling, might want to cluster first, then the user can label clusters as opposed to individual instances. Cluster (e.g., by Gaussian mixture models), user provides semantic labels.
See variational Dirichlet process (VDP) for learning the GMM (does not require a prior on the number of clusters). Other clustering algorithms such as density-based spatial clustering of applications with noise (DBSCAN) may also be appropriate.
IDEA: Just perform PCA over the 24x24 normalized images (raw feature vector), and cluster in this reduced space (4 or 5 dimensions). Could also think about applying locally linear embedding (LLE) instead of PCA. See \cite{Turk1991} and \cite{Belhumeur1997} for PCA. See \cite{deRitter2002} and \cite{Vanderplas2009} for discussion of LLE and classification.
Jeff Walls jmwalls@umich.edu