Thanks for your excellent work! How can I train and test on my own dataset use Dave?

[jerpelhan]

To train DAVE on your own dataset, follow these steps:

1. Train LOCA: First, train the LOCA model using the instructions provided on the GitHub here with your data.
2. Run train_det.sh: Use the trained LOCA model and run train_det.sh, updating the dataloaders to accommodate your dataset. The verification network configuration can likely remain unchanged.

Could you provide more details about your dataset? Specifically, is it significantly different from the dataset used in the original DAVE training?

[jerpelhan]

Yes, that's a different task. In FSC, the goal is to count all objects of the same class (e.g., pencils), not by color. I recommend running DAVE as usual with exemplars (including multiple colors of pencils) and then tweaking the verification (clustering) part. Instead of clustering based on exemplars, try different thresholds (similarity score) to see if it clusters objects by color. If the default version doesn’t manage to do this successfully, you may need to retrain the verification part of DAVE (feature projection).