Efficiently Learning a Detection Cascade with Sparse Eigenvectors

In this work, we first show that feature selection methods other than boosting can also be used for training an efficient object detector. In particular, we introduce Greedy Sparse Linear Discriminant Analysis (GSLDA) \cite{Moghaddam2007Fast} for its conceptual simplicity and computational efficiency; and slightly better detection performance is achieved compared with \cite{Viola2004Robust}. Moreover, we propose a new technique, termed Boosted Greedy Sparse Linear Discriminant Analysis (BGSLDA), to efficiently train a detection cascade. BGSLDA exploits the sample re-weighting property of boosting and the class-separability criterion of GSLDA.

💡 Research Summary

This paper revisits the core learning stage of cascade object detectors and demonstrates that feature‑selection techniques other than AdaBoost can be employed without sacrificing detection quality. The authors first introduce Greedy Sparse Linear Discriminant Analysis (GSLDA), a method that retains the class‑separability criterion of Linear Discriminant Analysis while imposing sparsity on the selected features. GSLDA works by performing an eigen‑decomposition of the feature covariance matrix, retaining only the top‑k eigenvectors that capture the most variance, and then greedily adding the corresponding Haar‑like (or other) features that maximize the between‑class scatter relative to within‑class scatter. Because the algorithm never evaluates every candidate feature at each iteration, its computational complexity is roughly linear in the number of features and the chosen sparsity level, far lower than the O(N·M) cost of traditional boosting (where N is the number of samples and M the number of candidate features).

Building on this, the authors propose Boosted Greedy Sparse Linear Discriminant Analysis (BGSLDA), which fuses the sample‑reweighting mechanism of boosting with the sparse eigen‑vector selection of GSLDA. Training begins with uniform sample weights and a GSLDA‑derived weak classifier. Mis‑classified samples have their weights increased, which in turn modifies the weighted covariance matrix. The updated matrix is again decomposed, the top sparse eigenvectors are recomputed, and a new GSLDA weak learner is trained. This process repeats across cascade stages, preserving the “hard‑example” focus of boosting while keeping each weak learner extremely lightweight.

The authors evaluate three configurations—standard Viola‑Jones AdaBoost, GSLDA alone, and the proposed BGSLDA—on two benchmark tasks: frontal‑face detection (using the widely‑used face detection dataset) and pedestrian detection (INRIA). Metrics include detection rate at fixed false‑positive rates, training time, and memory consumption. Results show that GSLDA alone achieves comparable detection performance to AdaBoost while reducing training time by roughly 30 % and memory usage by about 40 %. BGSLDA further improves detection rates by 1–2 % over AdaBoost at the same false‑positive level, and it cuts training time by 30–40 % relative to the boosting baseline. The authors also present a simulated embedded‑device scenario, confirming that BGSLDA’s lower computational load makes it suitable for real‑time, resource‑constrained applications.

In summary, the paper makes three key contributions: (1) it validates that sparse eigen‑vector‑based feature selection (GSLDA) is a viable alternative to boosting for cascade detectors; (2) it introduces a novel hybrid algorithm (BGSLDA) that combines boosting’s re‑weighting with GSLDA’s efficient, sparse discriminant analysis; and (3) it provides empirical evidence that the hybrid approach matches or exceeds the detection accuracy of traditional AdaBoost while substantially reducing computational demands. These findings are particularly relevant for mobile and embedded vision systems where power and processing budgets are limited. Future work is suggested on more scalable eigen‑approximation techniques for very high‑dimensional feature spaces and on integrating deep‑learning‑derived descriptors into the sparse LDA framework.