ArcGIS Pro 3.5.4 deep learning model building

[MarkSchweder]

Hi, I'm using ArcGIS Pro 3.5.4 to build a deep learning model to recognize specific birds in orthomosaic images. Brown pelicans are the subject at the moment. A model trained with 50 samples produced the blue bounding boxes when it was tested on the attached orthomosaic image. A supplemental training of the new model using 50 additional different samples (fine tuning) resulted in an even newer model that produced the red bounding boxes when it was tested on the same orthomosaic image. It doesn't appear things are improving. Do you have a suggestion to reverse that trend?

Thanks, Mark