I am developing a Body Measurement extraction application, my current stage is able to extract the point clouds of human body in a standing posture, from every angles.
Now, to be able to recognize shoulders, neck point etc, my research seems to fall into following flows:
- Obtain a lot of data, with labeled landmark points (shoulder left, shoulder right, neck line).
- Use PointCNN / PointNet++ to perform segmentation for each landmark.
- Once the landmarks are extracted, use Open3D / Point Cloud Library convex hull to obtain the measurement along point clouds.
- Method A seems straight forward, but might depends on the quality of point cloud, especially 3rd step.
- Obtain a lot of data, with labeled landmarks and also measurement of shoulder length, chest circumference etc.
- We first train the network to identify landmarks,
- Then from landmark, we record the distance to the next nearest point, train the network to obtain the measurement that we want.
- Obtain a lot of data, with measurement of shoulder length, chest circumference etc ONLY.
- Pick a random point, and we record the distance to the next nearest point, train the network to obtain the measurement that we want.
- How much data needed for this kind of learning?
- If I obtain training data from somewhere online, and later validate using my own scanned data, will that valid?
- Which method makes more sense?
- Which existing problem with solution is similar to my case? (Facial recognition?) That I can refer to it to solve my problem.
This is technically my first machine learning project, so please bear with me if my questions seems too silly.