Whether to annotate certain variations of my objects

I am annotating images of deer.

I have some deer that are lying down in the same image as standing deer. For those lying down, you clearly see the head and body, but not the legs. I feel like this is a huge digression from standing deer so I’m hesitant to label them as deer. They also will not be in play for my eventual use-case (spotting deer while driving.)

So, how would you recommend I handle this? Do I just not mark the lying deer but annotate the others? Should I delete any image with a deer lying down to avoid confusing the model?

Hi @Automatez - label the images as you would like your model to see them in practice (deployment).

If you don’t want the model to recognize any other deer as “deer”, save for the ones that are upright, then only label the upright deer. Just know you’ll need more examples of deer-imagery to get the model performing to your expectations, as it will first be confused as to why other objects that look highly similar are not labeled/should not be considered as “deer.”

Active Learning is a good process to help mitigate these issues of false detections much faster:

  1. What is Active Learning?
  2. Active Learning - Roboflow | pip install roboflow: The Computer Vision Python Package

Searching Roboflow Universe for “deer” or “deer object detection” should help with collecting more images, and also sampling from YouTube:

I don’t mind the model recognizing lying down deer. I was just worried that it would mess things up because suddenly these “shapes” don’t have 4 legs sticking down from them. But it sounds like you feel they would be a close enough match that I can include them and not confuse the model so I’ll do that.

Yup, you should be able to include. The model will have issues with false detections, more than likely, as you start out. But with an Active Learning approach, and retraining, it should begin to learn the difference.

  • (Active Learning links referenced in my previous replies)