Phone Camera (keypoint detection)

🤝 Community Help
1

Hello,

I am new to computer vision and I am trying to figure things out. Please I have a project that requires the recognition of various parts of a goat. I plan to get enough videos and pictures from goat markets. I have been learning with objects (trousers, spoons, leaves). I noticed that after annotation and trainning, the object of interest is recognized but the parts (keypoints are highly distorted).

1. Please, What can be the issue?
2. Is it okay to use a phone camera (i have been using a 50MP itel p55 camera)?
3. About how much pictures do i need to annotate for the computer to recognize the keypoints?
4. Do i need a plain background with same light condition?

2

Hey! Good starter project. Here are my thoughts on your questions.

  1. The issue could be so many things! Probably the biggest is having a consistently mapped dataset. That is, always marking a hoof exactly where the hoof is, and even including it if it’s hiding behind another hoof. And pick obvious points like hoof and nose. Trying to map “middle of their back” is going to be a mess. AI had a couple good checks as follows:
  • Left/right ambiguity
    Decide whether “left” means the goat’s anatomical left, not the viewer’s left, and label that way every time.
  • Side-view bias
    If most images are side views, the model may struggle on front/back/angled goats. Either limit the use case to side views or intentionally collect varied angles.
  • Occluded legs
    Four-legged animals often hide one leg behind another. Use occluded/deleted keypoint labels correctly instead of guessing invisible points.
  • Point choice
    Pick landmarks that annotators can identify reliably. “Shoulder” on a furry goat may be harder than “front hoof” or “tail base.”
  1. Phone camera is totally fine
  2. You could start at 50 and then train a model and see how it does. That should be enough to start
  3. The background and lighting just needs to be consistent with whatever you will need to use it for. So if it will only ever be finding keypoints on a goat in a barn, every photo you annotate can be a pic in a barn. But if you’ll want to drop in some pics of a goat in a field, you will want those in your dataset from the start.

Best of luck!

3

Thank you very much for the detailed response. I am grateful. I would be going back to take new pictures and to start a new labelling strictly following your guide.