search

neptune-ai / open-solution-mapping-challenge

Open solution to the Mapping Challenge :earth_americas:
https://www.crowdai.org/challenges/mapping-challenge
MIT License
377 stars 96 forks source link

Transfer learning using the available weights #234

Closed Ahsanr312 closed 3 years ago

Ahsanr312 commented 3 years ago

Suggest an idea for this project Hi! I have been following the repository and used the available weights for prediction on my own dataset which resulted in poor segmentation. While thinking about how to achieve better results on my own dataset, I came up with a few ideas as follows:

  1. Train the architecture only on my custom dataset
  2. Train the architecture on MS COCO + custom dataset
  3. Train the architecture using transfer learning on custom dataset

I believe, I won't get good results by choosing first approach whereas the second approach can get me some good results however I am looking for any method for the third approach mentioned above.

Looking forward for hearing from you all.