Reproduce Performance Discussion

Question

Reproduce Performance Discussion

khw321 opened this issue 6 years ago · comments

khw321 commented 6 years ago

Thanks for the nice job.
I run the eval_market1501.sh, and get such result,

And the train_market1501.sh result is

How can I get the result in your paper?
Thanks.

xxx2974 · Answer 1 · Fri Jan 11 2019 14:24:54 GMT+0800 (China Standard Time)

@khw321 Hi ，how can you get the pretrained model,I can't open the link "https://www.dropbox.com/s/2ljm35ztj6hllcu/bvlc_googlenet.caffemodel.pth?dl=0" and "https://www.dropbox.com/s/pzb3ow1793yf8dc/pose_iter_440000.caffemodel.pth?dl=0"

khw321 · Answer 2 · Fri Jan 11 2019 14:30:57 GMT+0800 (China Standard Time)

@xxx2974 You need to use a proxy.

yuminsuh · Answer 3 · Mon Jan 14 2019 17:12:09 GMT+0800 (China Standard Time)

Hi, thanks for your interest in our work.
The result of the code is the one without re-ranking. The corresponding numbers reported in the paper are
top-1 / top-5 / top-10
91.7% / 96.9% / 98.1%.
To get the result with re-ranking, I used the code on top of the above result.

The results in the paper are obtained with caffe. To reproduce the exact numbers, please use this model and weights trained on the market-1501 dataset. It needs custom caffe.