I want to recreate my hands in 3d in order to create a synthetic dataset of bone keypoints so I can train an ai to detect them on my actual hand in a webcam. I can already do the ai stuff on other datasets but i want one that is perfectly labeled.
What free software should I be using to scan my hand and what pitfalls should I be looking out for?
https://www.kaggle.com/code/ollielearnscode/hand-keypoint-26k-keras-claude
I've been trying to get gemini to teach me how to train keypoint predictors. I've tried with celeba and also with this hand keypoint dataset but the same problems is occurring- the predictions are just hovering around the average positions. I must be making some newbie error. Does anyone have a minute to help me out?
I've been using the celebA 5-keypoint dataset and my results have been that the markers are usually predicting the average location and not tracking well, particularly if the head looks to the side. Claude tells me this is likely becasue the dataset is centred on the face with most pointing forward. notebook here
I was wondering if someone could point me to a better challenge. My ultimate goal is to make a mocap system for myself. I'm looking for keypoint regression. doesn't have to be humans
or if i'm going about this wrong pls let me know
After reading https://microsoft.github.io/DenseLandmarks/ i want to have a go myself. It's been a few years since i tried any ML related stuff but i'm getting back up to speed.
Before doing the whole high density mesh, my plan is to start of with the 5 point eyes/nose/mouthcorners celeba dataset and then to make my own.
I have just about enough blender skills to make a human generator but i expect this to be the hardest part of the project.
Do you think I should try to train on mesh point prediction like the microsoft paper or perhaps train it on rig values?
What pretrained network should I use? I can't see any additions to the image networks in the past few years and it looks like mobilenetv3 would be a good one to use. Is it still in the realms on 224x224 networks?
It's a magic flying ship scenario where the trainee is failing and doesn;t eat dinner but then they don;t get ill and have to save everyone.
I'm struggling to think of ten minute stories that aren't blitheringly obvious