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
687 stars 159 forks source link

Reproduce resutls on Caltech dataset #153

Closed dogdogpee closed 1 year ago

dogdogpee commented 1 year ago

Claim: Cascade Mask R-CNN, Caltech dataset, 1.7%, 25%, 14 epoch

Reproduce: following configs/elephant/caltech)/cascade_hrnet.py/ got 10%. 58% on caltech at 14 epoch.

Question: what hyperparameters do you use to produce the claimed results?

Many thanks!

hasanirtiza commented 1 year ago

Refer to section 7 of our paper. It is about progressive finetuning and how we obtained this particular result. The hyperparameters are the one provided in the config file that you are using.

dogdogpee commented 1 year ago

Thanks for your reply!

Could you please share the training config of the Cascade R-CNN (HR Backbone) in Table 5 under the case of Caltech-->Caltech in the paper you mentioned?

Thanks for your time!

hasanirtiza commented 1 year ago

Did you try with the config provided in the repo ? I mean did you evaluate every checkpoint ? I do not exactly remember if it was the same config for table 5 and unfortunately I do not have the access to old files anymore (not working there anymore). Nonetheless, what I can suggest is, first start with the config provided in the repo, and in case required, just adjust the learning rate.