Problem: SVM training is expensive – Mining for hard negatives, bootstrapping Solution: LDA (Linear Discriminant Analysis). – Extremely fast training, very similar performance Claim
Linear Discriminant Analysis (LDA) Assumptions Learning - Classification
Implementation Features a simple procedure that allows us to learn a and a (corresponding to the background) once, and then reuse it for every window size N and for every object category.
Implementation Mean Covariance
Regularization Very large In my experiments 10, for making sure that is PSD.
Covariance
Fast training using LDA
Use in clustering
Clustering in WHO Space
HOG WHO
Clustering in WHO Space HOG WHO
(a) SVM Pedestrian Detection Linear Discriminant Models
SVM LDA Cen Pedestrian Detection Linear Discriminant Models
Results
MethodMean APTrain complexity Test complexity ESVM + Co-occ22.6High ESVM + Calibr19.8High ELDA + Calibr19.1LowHigh Ours full21.0Low
Results
Pascal NN Classification
Summary Whitened for HOG is better than HOG LDA for fast training of hog templates – Object Independent Background (?) mean better represents the cluster compared to the medoid – Use all the samples rather than 1 Their statistical models also suggest that natural image statistics, largely ignored in the field of object detection, are worth (re)visiting.