Notes about the data:
We did not use some tiles from the following full images because they included California Sea Lions in addition to Northern Elephant Seals:
4MSL0107, 4MSL0118, 4MSL0089, 4MSL0093, 4MSL0094, 4MSL0095, 5MSL0009, 5MSL0010, 5MSL0011, 5MSL0012, 5MSL0013, 5MSL0014, 5MSL0048, 5MSL0067, 5MSL0068, 5MSL0084, 5MSL0085, 5MSL3902, and 5MSL3903.
To reproduce our work...
- start up a docker container using the
Dockerfilein this repo. - tile the 230 images found in
Data/Full_ImagesusingCode/tile_images.ipynb - label tiled images using Roboflow
- train the model and evaluate it on the test set using
Code/run_model.ipynb - examine model errors using
Code/NOTEBOOK
Our final model weights are available here.