Hello everyone!

I am a freshman at the university. Lately, I have been interested in ML and DL approaches to solving problems. I want to build an image labeling/captioning model in a minor language. I have found that the language I am interested in has no labeled dataset.

I have three approaches in mind:

• Create dataset by myself - approximately 10000 images with manual captions - Decide the NN architecture train the model
• Try to use the existing pre-trained model and use the dataset I prepared
• Add Neural Machine Translation component in the architecture - For Multilingual captioning?

If possible, maybe I can cross-validate all these three options to see which one is potentially a better.

I am still learning and there are lots of unclear things I want to get some advice from the experts. Any insight or suggestion would mean the world to me!