baseline

[yja1]

i also try this, the result is very bad.

do you try table 5 in paper? I try, but the result of table 5 the first line only 68, far away to 87.5....

[levyfan]

I haven't tried that yet.

[levyfan]

I am confused with the difference between "Branch" and "Single" in Table 5. What do they mean?

[kilianyp]

Hi,
I also got confused by this. One interpretation could be:
Branch: Train on all branches, but only use one branch to evaluate.
Single: Train only with one single branch.

But not really sure.

[aishawn]

Why the result is very bad? Is this the same models as the paper described?

[levyfan]

Hi @aishawn , it is the same model. I can not re-produce what the paper claims.

[aishawn]

I find out the branch2 and branch3 256 channels part loss hard to converge. So I try only train part loss fisrt, it can reach mAP 0.8185 by this code. but still far away from 0.869..... and I think how to converge the global and part loss together is a problem, do you have some good ideas?

[SZULH]

hi,
1、I found that you should modified your code line 23 in mgn.py for python 3.
2、I run your code, but why the result is very bad? Isn't the training strategy different?
3、have you try other datasets on this model? eg: cuhk03、duke...

[levyfan]

Hi @SZULH,

Q1: welcome PR.

Q2: I am not sure. I jest feel that if you are working on your own papers, this paper probably is not a good start point (baseline).

Q3: No I haven't.

[rocketbear]

Hi @levyfan

I have been trying this method recently. I modify the code based on Person_reID_baseline_pytorch.

Currently I have only implemented the Part-3 (single) baseline as in Table 5 of the paper. The model is still training, and the Rank-1 accuracy on Market-1501 reaches 93.2% after 42 epochs, which is slightly higher than reported in the paper because I have used more data augmentation. The performance is also better than the PCB baseline in Person_reID_baseline_pytorch, which could be attributed to the z_g branch in the Part-3 (single) model.

It seems the results reported in the paper is promising and worthy trying.

[levyfan]

@rocketbear glad to hear! Can you share your code? I only get r@1=0.565321 on P3 single branch.

[Mr-zhu0912]

@rocketbear Hi,I have encountered the same problem, running the source code and not having the effect mentioned in the paper.Would you share your source code.Thanks a lot. My mail address is Zhuzhengfa0912@163.com