search

ZhihengCV / Bayesian-Crowd-Counting

Official Implement of ICCV 2019 oral paper Bayesian Loss for Crowd Count Estimation with Point Supervision
324 stars 83 forks source link

ShanghaiTech Dataset #33

Open hdasd opened 3 years ago

hdasd commented 3 years ago

Hi，could you tell me how topre-Process the ShanghaiTech Dataset?thinks!

enric1994 commented 3 years ago

I think preprocessing is not required for ShanghaiTech since the images are smaller and more uniform. You can run the training step directly You'll probably have to apply some changes to the proposed dataloader crowd_sh

12z34x commented 3 years ago

I think preprocessing is not required for ShanghaiTech since the images are smaller and more uniform. You can run the training step directly You'll probably have to apply some changes to the proposed dataloader crowd_sh

I do not agree with you. The min resolution of ShanghaiTech part A <256 and the min resolution of ShanghaiTech part B <512. So you will meet crop size error if you not use preprocess code. And I think the preprocess code is sane as QNRF. But this is a trick to use 512 min resolution "preprocess code" for 256 crop size of ShanghaiTech part A.(Because you enlarge the resolution)