How can I place more emphasis on low-level similarity and reduce the focus on semantic similarity?

[Tonniia]

Hi,

I’ve been collecting in-the-wild data pairs on my own. For the pairs shown below, where the same subject is captured from different angles and there’s no need to focus heavily on the semantic level, I’ve noticed some matching failures. Could you suggest which hyperparameters I should adjust to improve the results? (I have tried decrease t_init from 261 to 1, still fail)

Here is the data pair:

Thank you in advance for your guidance!