search

hasanirtiza / Pedestron

[Pedestron] Generalizable Pedestrian Detection: The Elephant In The Room. @ CVPR2021
https://openaccess.thecvf.com/content/CVPR2021/papers/Hasan_Generalizable_Pedestrian_Detection_The_Elephant_in_the_Room_CVPR_2021_paper.pdf
Apache License 2.0
689 stars 157 forks source link

Classes for the CityPersons dataset #112

Closed supremePSS closed 3 years ago

supremePSS commented 3 years ago

Hi, i have met some problems about the CityPersons dataset. There are 5 classes in the CityPersons, i.e. 'pedestrians', 'riders', 'sitting persons', 'other persons with unusual postures', and 'group of people'. I want to know whether your models were trained to predict the 5 class respectively, or just see all the objects as one class. I find your code in test_city_person.py writes all the predicted box as 'category_id'=1, shown as following: 1632915116(1) It seems like you see all the objects as one calss, is that right? thanks a lot!

hasanirtiza commented 3 years ago

No, that is not right, the code only treats pedestrians as valid samples as per the evaluation protocol. So during training, only ones with pedestrian class are used. The evaluation protocol in pedestrian detection literature(cityperson, ecp etc.), only considers pedestrians, all other classes are ignored.