xuyanwu / ral-net Goto Github PK

Hi,

Since you have not provided your pre-trained models, so we re-trained all the models again using the following code:

python RAL_Net.py --data-root='your data root' --epochs=10 --batch-size=512 --n-pairs=5000000 --loss-type=RAL_loss --lr=10 --augmentation=True

The results we obtained are different than the one reported in the paper (shown in brakets "( )" ).

Training on Liberty:
FPR95 on Notredame = 0.414 --(0.37)
FPR95 on Yosmite = 1.5879 -- (1.52)

Training on Notredame :
FPR95 on Liberty= 1.248 -- (1.30)
FPR95 on Yosmite = 1.336 -- (1.31)

Training on Yosmite :
FPR95 on Notredame = 0.706 -- (0.67)
FPR95 on Liberty = 2.558 -- (2.39)

Can you please provide the pre-trained models.
Also, which model do you use to evaluate on HPatches dataset.

Thanks in Advance.

Hello,thank you for your sharing!
I trained the model on liberty and tested on notredame successfully. But the following bug appears when i try to create test_loaders with yosemite dataset :

RuntimeError: $ Torch: invalid memory size -- maybe an overflow? at ..\aten\src\TH\THGeneral.cpp:188