I’m creating a dataset for detecting eliphant,Rhino in thermal images. on my first trial of model the model was detecting eliphant like sign board or eliphant like any other pattern as eliphant , so for avoiding it i have taken those images and while annotation i have marked them as null (explicitely) , and after training again i’m getting same issue , what am I doing wrong ?
I have read the below blog also for differences
should i mark it null explicitely while annotating or just leave those images so that my model knows to avoid those images.
box F1 score all classes 0.96 at .41 conf
box PR score all classes 0.98 at map@ 0.5
box recall confidence curve 1 at 0.0
box precision all class 1 at 0.93
Label Distribution is a bit imbalanced.
Rhinos (~2500 instances) vs Elephants (~1000 instances).
You mentioned that you’re attempting to detect objects in thermal images. Are all the images in the training and validation sets thermal images? …including the background images?
The metrics you’re showing appear to be quite decent, so it makes me think that something else is awry with the data. It might be helpful to share some example images from training and testing (where you observe the false positives).
