Article directory
1. Four questions
1. What problem to solve
The current self-reconstruction mainly learns global geometry and fails to effectively explore local geometry. Exploring unsupervised learning of local geometry and lack ofEfficient and semantic local structure supervision signals
Because of lacking effective and semantic local structure supervision, however, error may accumulate in the local structure learning process, which limits the network’s ability in 3D point cloud understanding.
2. What is the solution
Three branches: described in detail in 3. Overview
- (General) branch A: Aggregate all features with RNN
- (above) branch R: Rebuild monolith, global, with VAE? Can generate more feature space
Self-reconstruction is started from a variational feature space, which enables MAP-VAE to generate new shapes by capturing the distribution information over training point clouds in the feature space
- (below) branch P: split the original point cloud input into the front half and the back half, half-to-half prediction
We introduce multi-angle analysis for point clouds to mine effective local self-supervision, and combine it with global self-supervision under a variational constraint.
3. How does it work?
The effect on ModelNet:
segmentation on Shape part dataset (mainly compared to LGAN? Haven't seen it...)
4. What problems still exist
?
2. Introduction to the thesis
3. References
4. Harvest
Research motivation? : Intuitively, by observing the point cloud from different angles, the correspondence and relationship between different shape regions can be clearly displayed, that is, the correspondence between the front half and the back half of the shape in each view
Innovation:?
- multi-angle: The division of angles is also very important, I don't see much
- variational constraint: VAE? Variational encoder? Ability to generate more shapes (what VAE does)
A new proxy task is provided: 1. Divide half and predict half; 2. Rebuild the whole (common)
Keywords: semantic, local structure