system123
commented
3 years ago 1) The code to train the goodness networks is included. However, after training the correspondence network you need to use the trained model to generate the features for training the goodness network. I plan to fix the readme to include these details in the near future – I am currently no longer at TUM and thus this project is not my main focus anymore, so please be patient. 2) There are many applications of the located corresponding points – one of which is to fix optical geocoding. In this case we expect the optical and SAR imagery during testing to be somewhat aligned (i.e. the optical patch should be somewhere within the SAR search window). The larger you make the SAR search window the more likely the matching is to fail (as there is more scope for false positives). So ideally you'd want the SAR search window to be large enough to account for the maximum expected offset, but small enough that the search area is limited.
Sorry to bother you again. Could you please upload the code for the training of the goodness network?(since we only got the code for generating the heatmap) Another question: is the finnal aim of the framework to correct the geocoding of the optical image? so during testing procedure, the optical and the sar image is roughly matched at the beginning?