search

huanghoujing / beyond-part-models

PCB of paper: Beyond Part Models: Person Retrieval with Refined Part Pooling, using Pytorch
331 stars 81 forks source link

Testing results on CUHK03 #6

Closed iodicesara closed 6 years ago

iodicesara commented 6 years ago

Hi, I tried to train the model on CUHK03 and repeat the experiment, but results I got are lower compared to those reported on the paper, i.e. Single Query: [mAP: 41.74%], [cmc1: 46.36%], [cmc5: 67.43%], [cmc10: 75.71%] Re-ranked Single Query: [mAP: 58.11%], [cmc1: 56.71%], [cmc5: 69.71%], [cmc10: 77.00%]

Do you know any reasons of that? Thanks.

huanghoujing commented 6 years ago

Well, maybe it's because of some little but non-trivial differences between my implementation and the author's, which are not detailed in the paper. I have no idea also : )

huanghoujing commented 6 years ago

The PCB baseline on CUHK03 is now quite normal after using independent 1x1 conv for each part. On CUHK03, the Rank-1 accuracy and mAP are about 2 and 0.5 points lower than the paper, respectively.