Closed aruba01 closed 2 years ago
That is very strange behaviour that I have not seen before. Having a mismatch between the network and the scene might be one explanation. May I ask how you call exactly to get those results?
I can reproduce your results. There is a mixup in the filenames of the pre-trained models. The following combinations yield the correct results:
python test.py Cambridge_OldHospital models/rgbm/cambridge/Cambridge_ShopFacade.net --mode 1
python test.py Cambridge_ShopFacade models/rgbm/cambridge/Cambridge_StMarysChurch.net --mode 1
python test.py Cambridge_StMarysChurch models/rgbm/cambridge/Cambridge_OldHospital.net --mode 1
I will try to trigger an update of the models data package. Thanks for brining it up!
I updated the data package of pre-trained models to fix the naming error. Everything should work now.
I test the pre-trained network on cambridge datasets. The results of OldHospital, ShopFacade and StMarysChurch are obviously deviated from the normal value, while the results of KingsCollege and GreatCourt are normal. I guess that the pre-trained network do not correspond to dataset. The following is the screenshot of ShopFacade: