search

blue-blue272 / VideoReID-TCLNet

Code of Temporal Complementary Learning for Video Person Re-Identification (ECCV2020)
67 stars 19 forks source link

How to get the paper's result on Mars dataset? #2

Closed AsuradaYuci closed 4 years ago

AsuradaYuci commented 4 years ago

I run the code with the default setting, but get the best results: Extracted features for query set, obtained torch.Size([1980, 4096]) matrix Extracted features for gallery set, obtained torch.Size([11310, 4096]) matrix Extracting features complete in 1m 14s Computing distance matrix Computing CMC and mAP Results ---------- mAP: 80.83% CMC curve Rank-1 : 88.03% Rank-5 : 95.25% Rank-10 : 96.62% Rank-20 : 97.32%

while the paper's result are 85.1% mAP, 89.8%Rank1. The mAP is really high.

blue-blue272 commented 4 years ago

@AsuradaYuci You can run the code with one GPU, and try again.

And I think your results are the train results (with 4 frames). When testing, we use all the frames. You can run the test files with the model weights.

AsuradaYuci commented 4 years ago

@blue-blue272 Thanks, I run all the frames, and get a better results. But it cost about more than 1 hour to test...

blue-blue272 commented 4 years ago

@AsuradaYuci Yes. Usually, it costs about 2 hours for me to test a model. It's really annoying.