megvii-research / Iter-E2EDET

Official implementation of the paper "Progressive End-to-End Object Detection in Crowded Scenes"

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

quesions about dataset

jiangduwang opened this issue · comments

hi,author!
The data set link you provided includes three training sets, one test set and one verification set. How do you put these files under the CrowdHuman folder?
Thanks!

commented

Please check this for the detail.

Hi,
I have tried all train sets provided in CrowdHuman website but it keeps throwing errors for missing pics. And I also find very hard to identify the instruction in the file you mentioned. From my aspect, all train sets would be fine for just training step.
Could you please give us a clear clue to going on the experiment?
Thanks so much.

commented

The problem may be derived from the unfamiliarity with the detectron2. Here is the procedure to curate the CrowdHuman dataset for both training and testing.

  1. please download both training and validation datasets, and mix the separated training or validation dataset. Such as merge CrowdHuman_train01, CrowdHuman_train02 and CrowdHuman_train03 into one CrowdHuman_train dataset, so does the validation set.
  2. Change the directory file at this to ensure the project can find image and json files properly in your own system. Besides, this code describes how to register the CrowdHuman dataset in detectron2.
  3. If a loading image error still occurs after the above procedure is completed, this line decribe the way the project load an image. Your can come up with an approach to load image properly. Such as add another key/value in the train.json/ val.json to redirect the proper image file location in your disk.

Really appreciate your very kind and patient answer, and it's quite clear now!
Thanks again for your help.