Why the Det and Emb test are quite different from those in papers?

[guangyaooo]

Has anyone tested AP and TPR for the pre-trained model jde.uncertainty.pt? The results of my test are AP=80.54 and TPR@FAR =84, which are quite different from those of AP=83.0 and TPR=90.4 claimed in the paper. Except the batch_size=8 All other hyperparameters were set as default. The GPU used for this result is TITAN V. Part of the output for Detection test:

Part of the output for Embeding test:

[salt-fisher]

I have tested the mAP from the released pre-trained model.(Notice that the test dataset in configuration needs to be modified, the test dataset in paper only use Caltech). And i get a 0.8247 mAP, still a little lower than the 0.83 in paper. Beside, the released pre-trained model seems to train 60 epochs, while released code trains for 30 epochs

[guangyaooo]

I have tested the mAP from the released pre-trained model.(Notice that the test dataset in configuration needs to be modified, the test dataset in paper only use Caltech). And i get a 0.8247 mAP, still a little lower than the 0.83 in paper. Beside, the released pre-trained model seems to train 60 epochs, while released code trains for 30 epochs

Thank you very much, I'll try it again later. Did you test the TPR score for jde.uncertainty.pt?

[Qizec]

I have tested the mAP from the released pre-trained model.(Notice that the test dataset in configuration needs to be modified, the test dataset in paper only use Caltech). And i get a 0.8247 mAP, still a little lower than the 0.83 in paper. Beside, the released pre-trained model seems to train 60 epochs, while released code trains for 30 epochs

Thank you very much, I'll try it again later. Did you test the TPR score for jde.uncertainty.pt?

Excuse me ,What is your final TPR score and AP?thanks