How many epochs can achieve performance of 80% as demonstrated in the paper without a pre-trained model while training

OnderT / XoFTR

XoFTR is a cross-modal cross-view method for local feature matching between thermal infrared (TIR) and visible images.

Apache License 2.0

61 stars 0 forks source link

How many epochs can achieve performance of 80% as demonstrated in the paper without a pre-trained model while training #7

Open YanQinDeng opened 1 week ago

YanQinDeng commented 1 week ago

hello, Now I am reproducing your work. But training is very slow.How many epochs can achieve performance of 80% as demonstrated in the paper without a pre-trained model? Could you show me your training curve of AUC? I'm looking forward to your earliest reply.

OnderT commented 1 week ago

Hello,

We train our model for 30 epochs, which takes about 24 hours using 8 A100 GPUs. Below, I’ve shared the AUC and loss curves from a training run with pre-trained weights. When starting with untrained weights, the curves look similar, but the scores change slightly.

Note: The AUC curves are based on the validation set, not the benchmark sets. The validation set doesn't distinguish between sunny and cloudy conditions, unlike the benchmark sets. xoftr_auc