Feature-centric evaluation for efficient cascaded object detection

We describe a cascaded method for object detection. This approach uses a novel organization of the first cascade stage called "feature-centric" evaluation which re-uses feature evaluations across multiple candidate windows. We minimize the cost of this evaluation through several simplifications: (1) localized lighting normalization, (2) representation of the classifier as an additive model and (3) discrete-valued features. Such a method also incorporates a unique feature representation. The early stages in the cascade use simple fast feature evaluations and the later stages use more complex discriminative features. In particular, we propose features based on sparse coding and ordinal relationships among filter responses. This combination of cascaded feature-centric evaluation with features of increasing complexity achieves both computational efficiency and accuracy. We describe object detection experiments on ten objects including faces and automobiles. These results include 97% recognition at equal error rate on the UIUC image database for car detection.