Summary
Use Max Pooling to get unordered global feature. If per-point segmentation is needed, concat it with local point feature and voila!
Properties of Point Sets in
- Unordered. It doesn’t matter which point comes first. PointNet use a big max pooling to make this happen.
- Interaction among points. It need to capture local structures from nearby points. I don’t see this part from PointNet.
- Invariance under transformations. Rotate the points and the result is still the same. Accomplished by having a mini PointNet (T-Net) output a rotational matrix, thus align them to a canonical space. Not much performance improvement though.
Model structure
Note that the dimension of the global pose feature greatly affect the performance of the model.
Intuitively, the network learns to summarize a shape by a sparse set of key points.