About train and evaluation of HFNet.

[haoweiz23]

Hi authors,

Thanks for your contributions. I have serveral questions about training and evaluation details of HFNet:

1. I followed the readme to export descriptors and localize poses. Then I submit the localization results on visuallocalization.net and only get correct results on Aachen dataset. Both RobotCar-Seasons and CMU Seasons datasets fail (robotcar day all: 13.9 / 21.1 / 25.3 and CMU Seasons urban: 0.0/ 0.0/ 0.0). It seems that I have submitted the imcomplete or wrong datas. Do you have any advices? By the way, I can only get day and night results of Robotcar on visuallocalization.net, how can I get the four robotcar scenes evalutation results sepereately which have represented in your paper?
2. Will we get the better or worse localize results after we finetune on target (Aachen / Robot / CMU) dataset? Is the training code the same as the finetune code?
3. Could you provide more details about your traning tricks? Such as the hardware enviromental, gpus, training times.