About the NWPU-Crowd dataset

[little-seasalt]

Dear author, I encountered some problems in the process of reproducing the NWPU data set: I screened the best models through the indicators of the validation set. After 1500 rounds of training, the indicators of the validation set reached about 73, and then I tested the test set on the best model, and submitted the results to the official website. I only got an MAE of 112. Do you know what the possible reasons are? Also, I would like to ask what is your approximate indicator on the validation set?Thank you for your answers in advance.

[cxliu0]

I think this issue is similar to the UCF-QNRF dataset. For large-scale datasets, it is important to ensure that the training patches contain a sufficient number of people. If many training patches are empty, the training supervision would be weak.

Regarding performance on the validation set, I remember the MAE is between 40 to 50.