Cascade object detection with deformable part models

A survey on face detection in the wild: Past, present and future

“…We will describe several approaches for reducing the complexity without significant performance degradation [165,166,167,168,169]. Finally, we discuss about a DPM model that produces state-of-the-art results in recent benchmarks.…”

“…Even the fully shared model needs many seconds per image, making it difficult to apply DPMs for real-time applications. Recently, efforts have been made to develop DPMs with decreased computational complexity achieving near realtime performance [165,166,167,168,169]. The main steps of DPMs that allow optimization include (a) feature extraction using HoGs, (b) filter design, (c) computation of filter responses (correlations) and (d) score computation (mainly using cascade methods).…”

A Real-Time Human Detection System for Video

“…With the detector output's guidance, a fast tracking by detection method is introduced by combining object association and meanshift tracking [24]. The final system runs 20 fps for the VGA resolution video, with a better accuracy than some state-ofthe-art methods [1], [8], [12], as well as more than 10 times speedup.…”

“…For PETS 2006 dataset, we choose S4-T5-A-4 (2 minutes long) which has adequate number of humans and is similar to our application scenario. Apart from HOG, two additional state-of-the-art public available human detectors' results are given as a comparison (boosted Histogram [12] and cascaded Deformable [8]). The result of our previous fast detector (30 fps on QVGA) with matrix based structure [28] is also given.…”

“…D.1979 gorithms, the most common ones are linear SVM and boosting. Detection accuracy can be improved by utilizing more complex features (such as covariance feature [6]), integrating heterogeneous features [4] or employing sophisticated learning algorithms (such as the intersection kernel SVM [7] or latent SVM [8]), but they suffer from greatly increased computation burden and are infeasible for real-time applications. Some studies have focused on improving the detection speed.…”