Where is matching probabilities and matching success calculated

facebookresearch / silk

SiLK (Simple Learned Keypoint) is a self-supervised deep learning keypoint model.

GNU General Public License v3.0

644 stars 58 forks source link

Where is matching probabilities and matching success calculated #47

Closed akimihiro closed 1 year ago

akimihiro commented 1 year ago

I am looking for specific code that implements the calculation of matching probabilities and matching success, but I'm having a hard time finding it. It seemed to me that the self._model.coordinate_mapping_composer.get function in SiLKRandomHomographies within lib/models/silk.py does it. However, upon examining the model.coordinate_mapping_composer.get function, I still cannot find the specific code where the calculation is implemented.

gleize commented 1 year ago

Hi @akimihiro,

The coordinate mapper doesn't handle the matching, it handles the coordinate change between the feature space and the input space (because the feature spatial dimensions are downsampled). The matching probabilities can be found two places :

During training, the matching probabilities are computed in log space and implemented here. Regarding the matching success, we first extract the best prediction of the model here and test if they are a success here.
During inference, the matching probabilities are computed here.

akimihiro commented 1 year ago

Thank you very much for your response! It successfully resolved my issue. I appreciate the time and effort you took to help me out. Your clear explanation made it easy for me to understand the code. Thanks again!