Open ducha-aiki opened 1 year ago
Hi,
yes this seems to be an issue here. You can fix it with "data.shuffle_training=True" in glue-factory. This might be an artifact from code refactoring, we will fix this.
Did not quite catch the issue here. Does this issue still exist in glue-factory codebase (since this is the LightGlue codebase). During my test running "data.shuffle_training=True" produces identical results as the default settings. @Phil26AT
Hi, did you train on the entire Homography dataset? I used a subset of the Homography dataset for training and the results were particularly poor
Hi,
a bit of noob question - should I add some shuffling, or this is expected for homography pretraining?