Question about fine-tuning WeDetect-Tiny on English-annotated design datasets

[nzomi]

Thank you for your great contribution, and congratulations on WeDetect being accepted to CVPR 2026! 🎉

I'm very impressed by WeDetect's speed compared to other open-vocabulary detectors. I'm planning to fine-tune WeDetect-Tiny for a design-domain task: recognizing artistic text, graphics, logos, etc. on flat design images (like T-shirt prints, cushion covers, and tote bags), in order to separate editable elements from backgrounds.

However, I've noticed that most open-source datasets in the design field (e.g., InfoDet, ARText, Logo datasets, LICA) are annotated in English, while WeDetect was pre-trained with Chinese annotations. This makes me curious about the model's cross-language transfer learning capability.

Could you please share some suggestions or best practices for the fine-tuning stage in this scenario? For example:

Should I translate the English class labels into Chinese to better align with the pre-trained text encoder?

Or can WeDetect generalize well to English prompts directly?

Are there any recommended fine-tuning strategies (e.g., learning rate, PEFT vs full fine-tuning) for this kind of cross-lingual adaptation?

Any advice would be greatly appreciated. Thanks again for your amazing work!

[fushh]

Thank you for your interest!

1. It is recommended to translate the class names into Chinese to better align with the pretraining.
2. It appears you are working under a closed-set setting. In this case, you can fine-tune the model without relying on a language model. However, it is advisable to initialize the prompts with the corresponding class embeddings to improve performance.
3. Full fine-tuning generally yields better results and is therefore recommended.