It's Not All About Size: On the Role of Data Properties in Pedestrian Detection

Amir Rasouli, Iuliia Kotseruba, John K. Tsotsos; Proceedings of the European Conference on Computer Vision (ECCV) Workshops, 2018, pp. 0-0

Abstract


Pedestrian detection is central in applications such as autonomous driving. The performance of algorithms tailored to solve this problem has been extensively evaluated on benchmark datasets, such as Caltech, which do not adequately represent the diversity of traffic scenes. Consequently, the true performance of algorithms and their limitations in practice remain understudied.To this end, we conduct an empirical study using 7 classical and state-ofthe-art algorithms on the recently proposed JAAD dataset augmented with 16 additional labels for pedestrian attributes. Using this data we show that the relative performance of the algorithms varies depending on the properties of the training data.We analyze the contribution of weather conditions and pedestrian attributes to performance changes and examine the major sources of detection errors. Finally, we show that the diversity of the training data leads to better generalizability of the algorithms across different datasets even with a smaller number of samples.

Related Material


[pdf]
[bibtex]
@InProceedings{Rasouli_2018_ECCV_Workshops,
author = {Rasouli, Amir and Kotseruba, Iuliia and Tsotsos, John K.},
title = {It's Not All About Size: On the Role of Data Properties in Pedestrian Detection},
booktitle = {Proceedings of the European Conference on Computer Vision (ECCV) Workshops},
month = {September},
year = {2018}
}