Year: 2021
Type: article
Abstract: Since the surge in popularity of Histogram of Oriented Gradients (HOG) in 2005 as the de facto feature vector for pedestrian detection, there have been many improvements in the detection pipeline that enable state of the art performance to be applicable to many real world problems. Nonetheless, the datasets available for training models have many biases, making it hard to use to detect pedestrians from videos and images obtained from other sources than the datasets.\nThis article presents a protocol to evaluate how pedestrian models generalize between different datasets. The protocol roughly consists of training a model with each dataset or dataset combination, and evaluating with the remaining dataset in each case.\nWe use the protocol to evaluate the performance of a typical pedestrian classification model based on HOG and/or LBP features and a SVM classifier. Alternatively, we also use a modern ConvNets model, to verify that the results of the protocol are due to the datasets and not the model.\nWe evaluate the models with the three most used datasets for pedestrian classification: INRIA, Daimler and TUD-Brussels. Our results show that while each dataset presents real world scenes, there are significant biases in each dataset that prevent models trained on one dataset to generalize to other datasets. Models trained on two fused datasets perform only marginally better on the third dataset than models trained on individual datasets, both for SVM and ConvNet classifiers.
Institution: Universidad Nacional de La Plata
Language: Spanish
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W3169752128
