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.

[pk429]

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.

Does this mean that if I want to train a model that can recognize crowded crowds well, the training data images should try to keep the crowd dense enough to ensure that the patch contains a lot of features? I trained a model using the NWPU original dataset and found that for blurry and dense crowded crowds, point recognition is prone to produce a grid, which is obviously incorrect

[cxliu0]

Yes. If the model did not see crowded scenes during training, the model may not output good results in such scenarios during testing.