TY - JOUR
T1 - REBoost: Probabilistic resampling for boosted pedestrian detection
AU - Lai, Shiming
AU - Zhang, Maojun
AU - Liu, Yu
AU - Theobald, Barry-John
PY - 2011
Y1 - 2011
N2 - Cascaded object detectors have demonstrated great success in fast object detection, where image regions can quickly be rejected using a cascade of increasingly complex rejectors/detectors. Although such cascaded detectors typically are fast and require minimal computation, they usually require iterative training, where classifiers are retrained to optimize rejection thresholds after testing on a validation set. We propose a cascaded object detector that uses probabilistic resampling for boosting reweighting, which has the advantage that only a single training step is required. Decision thresholds can be tuned on a validation set without the need for classifier retraining. Empirical results on a pedestrian detection task demonstrate that this reweighting results in a strong classifier that quickly rejects image regions and offers higher accuracy than other competing approaches.
AB - Cascaded object detectors have demonstrated great success in fast object detection, where image regions can quickly be rejected using a cascade of increasingly complex rejectors/detectors. Although such cascaded detectors typically are fast and require minimal computation, they usually require iterative training, where classifiers are retrained to optimize rejection thresholds after testing on a validation set. We propose a cascaded object detector that uses probabilistic resampling for boosting reweighting, which has the advantage that only a single training step is required. Decision thresholds can be tuned on a validation set without the need for classifier retraining. Empirical results on a pedestrian detection task demonstrate that this reweighting results in a strong classifier that quickly rejects image regions and offers higher accuracy than other competing approaches.
U2 - 10.1117/1.3658762
DO - 10.1117/1.3658762
M3 - Article
SN - 1560-2303
VL - 50
JO - Optical Engineering (OE)
JF - Optical Engineering (OE)
IS - 12
M1 - 127203
ER -