OpenCV's default HOG-feature based pedestrian detector is shipped with a sub-optimal linear SVM model. An improved version trained through an optimized bootstrapping method is being made public here. The C++ source code can be used to evaluate the retrained models for both Linear and HIK SVM. Corresponding models trained on INRIA and Caltech datasets have been included. Some visual results are also shown. The retrained model gives 9% lower Miss Rate (MR) on INRIA dataset than generally reported in the literature. Please cite this work as follows.

M Bilal, MS Hanif, "Benchmark Revision for HOG-SVM Pedestrian Detector through Reinvigorated Training and Evaluation Methodologies", IEEE Transactions on Intelligent Transportation Systems, 2019.