k-fold Subsampling based Sequential Backward Feature Elimination

TL;DR

This work introduces a wrapper-filter hybrid feature selection method that combines $k$-fold subsampling with sequential backward elimination (SBE) in conjunction with a linear SVM for human detection. It defines a dual scoring scheme, including step performance $P$ and feature relevance $R$, augmented by a mutual-information counter score to mitigate over-pruning. The algorithm produces compact feature subsets (e.g., 40%–80% of the full feature set) that can speed up detection by over 50% while maintaining or slightly improving accuracy, and it yields appearance models built from the selected features that outperform baselines like LatSVM-V2 and approach or exceed deformable part model performance. Experiments on INRIA and ETH datasets using HOG features demonstrate both speedups and robust generalization, with the 10HOG-V2 appearance model achieving up to 9% higher detection rates than LatSVM-V2 root filters on INRIA. The method extends to broad data scenarios and does not require highly precise object annotations, enhancing practical applicability for pedestrian detection systems.

Abstract

We present a new wrapper feature selection algorithm for human detection. This algorithm is a hybrid feature selection approach combining the benefits of filter and wrapper methods. It allows the selection of an optimal feature vector that well represents the shapes of the subjects in the images. In detail, the proposed feature selection algorithm adopts the k-fold subsampling and sequential backward elimination approach, while the standard linear support vector machine (SVM) is used as the classifier for human detection. We apply the proposed algorithm to the publicly accessible INRIA and ETH pedestrian full image datasets with the PASCAL VOC evaluation criteria. Compared to other state of the arts algorithms, our feature selection based approach can improve the detection speed of the SVM classifier by over 50% with up to 2% better detection accuracy. Our algorithm also outperforms the equivalent systems introduced in the deformable part model approach with around 9% improvement in the detection accuracy.

Figures (6)

• Figure 1: Proposed Feature Selection System Overview.
• Figure 2: Performance illustration: (a) Step performance curve of the proposed feature selection algorithm. (b) Human Detection performance using different HOG detectors trained with different feature subsets. The evaluation was carried out on the first 100 INRIA Full Images. The computed miss rates between 0.01 and 1 false positives per-image (FPPI) are shown in the legend. HOG Detectors trained with optimal feature subsets perform better than or similar to 100HOG showing a small overfitting problem.
• Figure 3: Appearance models trained using the proposed feature selection algorithm. (a) INRIA Dataset, (b) MIT Dataset and (c) CVC04 Dataset CVC04. The detection performances of the INRIA appearance models are evaluated on the first 100 INRIA full images. The appearance model from 20HOG shows 10% better performance than the root filter of LatSVM-V1 LatSvm-V1.
• Figure 4: Filter Score Computation: The positive filter score is subtracted from the negative filter score.
• Figure 5: Localisation accuracy: The human localisation accuracy of the 10HOG-V2 Filter is compared to that of the root filter of the LatSvm-V2 Model LatSvm-V2. Top - Score maps of the 10HOG-V2 filter at three scales. Bottom - Score maps from the LatSvm-V2 root filter at the identical scale. The score map from the 10HOG-V2 filter shows less noise than the one of the LatSvm-V2 root filter.