sjtuzq / person_search_gcn Goto Github PK

It seems that the codes in this project are just parts of the total model. You extract features for gallery and probe, and load them as input data for the training of GCN model in this project, right?
So, why dont you train the model end-to-end?

Hello,
can you show us the code about prepare the data to get your predata and testdata? Cause I cannot find directly use prw dataset as your well prepared CUHK data feature with the code in Joint Detection and Identification Feature Learning for Person Search.
Hope you can reply me and thanks for your help.

Hi,

In the paper,
"During training, the cosine embedding loss is jointly optimized with OIM loss"

But I can only see the cosine embedding loss, I could not find the implementation of OIM loss.
Is it now implemented??

GCN's input is extracted from Joint Detection and Identification Feature Learning~(CVPR2017)'s article? You only release CUHK-SYSU's predata. Do PRW dataset need to be processed through that paper code and entered into Context Graph?

When I run python train_gcn.py, I got the error as blew:

File "train_gcn.py", line 50, in test_one_epoch
recall_rates = np.load(os.path.join(data_root,'recall.npy'))

Is there lost the recall.npy file?

Hi, Thanks for the great work.!

I have some questions.

I looked at into your code in the above image, but I think it is different to your paper explanation.

I thought your algorithm was dividing the bbox in horizontally into several parts, and feeding each part into convolution layer which output 2048-dim. Then, calculating cosine similarity between gallery and probe persons.

But I can't find the part where it divides into partition, it seems it is feeding the whole bbox. Am I correct? or did I misunderstood..

Thanks in advance.

Hi, thanks for your great work.

The code run failed due to missing recall.npy
Could you help me? I will be grateful.