Low Localization Performance on Carpet Category Using Provided generated_data

[yulimso]

First of all, thank you for this excellent work. The paper is truly insightful, and I’ve learned a lot from implementing the provided code.

While training the train-localization model with the generated_data offered by the authors, I observed that most categories achieved good performance during testing. However, the carpet category consistently shows significantly lower localization performance and cannot be properly reproduced.

Has anyone else encountered a similar issue with the carpet category?
Could this be due to specific characteristics of the generated data or something else I might be missing?

Any guidance or suggestions would be greatly appreciated.

Thank you!

(*For reference, the implementation works correctly when using the provided pretrained weights for the localization model.
The issue arises only when I train the localization model from scratch using the provided generated_data.)