search

HanjiangHu / DASGIL

Code and pretrained models for our TIP work "DASGIL: Domain Adaptation for Semantic and Geometric-aware Image-based Localization".
24 stars 7 forks source link

Question about the triplet loss #7

Closed shenyanqing1105 closed 1 year ago

shenyanqing1105 commented 2 years ago

In the process of running the train.py, the triplet loss will be 0， is it normal? image

HanjiangHu commented 2 years ago

Thanks for your interest in our work. It may be normal because the distance in the latent space is getting larger for both positive and negative pairs, and it does not influence the final results too much. Maybe you can modify it with other more advanced techniques in contrastive learning and welcome to create pull requests if you find better methods or hyperparameters.