Where can i find the hdf5 and json files for the models? about camera-trap-classifier-publication-code HOT 2 CLOSED

So i managed to find the models(https://conservancy.umn.edu/handle/11299/199818) but not having much luck getting it to identify animals in still pictures. Im testing out trail camera images of bobcats and it cant identify those at all

from camera-trap-classifier-publication-code.

Hi sanktanglia

Thanks for your feedback. The model trained on data from Snapshot Wisconsin unfortunately has not seen many Bobcats (roughly 300) and thus may not reliably find them when the model is applied on a different dataset. Regardless, there may be opportunities for you to exploit the model and/or machine learning:

1. Reliably find true negatives (e.g. you are interested in Bobcats, thus filter out Deer images or images without any animals).
2. Possibly investigate the model scores to find a threshold that works for you. Example: The model classifies image as Deer with 70% probability but has 20% probability on Bobcat. In that case manually investigate the image (confidence thresholding approach).
3. Re-train / Fine-tune a model (see this repository with newer models) on data from your dataset--it is a common issue that models don't perform well on new / unseen datasets without some re-training (transfer-learning).
4. If you are only interested in a specific target class you may be able to train a model that identifies this class versus the rest, that might be more efficient to train and apply.
5. There are other models out there that may help your case. For example: Tabak et al.

All the best!

from camera-trap-classifier-publication-code.