FILTR: Extracting Topological Features from Pretrained
3D Models
Topology and 3D Deep Learning
- 3D encoders are trained on semantic and geometric tasks
- To make models topology-aware, we explicitly incorporate
topological summaries along with the point cloud.
Computing topological descriptors is costly.
Contrast with the data-driven principle of modern deep learning.
Some Usecases
Robust segmentation
Massive scalar fields characterization
Physically plausible reconstruction
Contributions
1
Topology Analysis
Evaluating topological understanding of 3D models.
2
FILTR: Filtration Transformer
Feed-forward topology prediction from frozen encoder features.
3
DONUT: Dataset Of maNifold strUcTures
Benchmark with topological annotations.
Quantifying Topological Understanding
Transformer encoderself-supervised reconstruction
Reconstruction
Underlying topology
Quantifying Topological Understanding
Probing
Global structure understanding
Feature Space Similarity
Local/multiscale understanding
- 30K meshes
- Connected components and genus labels
Probing Results
Pretrained encoders have a limited understanding of the global structure
of point clouds.
Quantifying Topological Understanding
?
Feature space similarity
Local/multiscale understanding
What topological signature?
Persistent Homology: A Multiscale Topological Descriptor
TERMINOLOGY ALERT
- Filtration
- Persistence diagram
- The farther from the diagonal the more salient the corresponding feature
Feature Space Similarity
Point cloud
Persistence diagram
Encode
Vectorize
Centered kernel Alignment
Alignment Results
Encoder features correlate with
vectorized persistence
diagrams.
3D models might implicitly capture
local/multiscale topological structure.
FILTR: Filtration Transformer
End-to-End Object Detection with Transformers, Carion et al., ECCV 2020
Quantitative Results
16.02
Point-MAEC
39.80
PointGPTL
40.19
PointGPTL
Qualitative Results
Predicted diagrams recover the main structures across
ModelNet40, and ABC, although trained on DONUT
Predictions are approximate.
Most persistent features are not captured.
3 Key Contributions
1
Topology Analysis
3D models capture local topological information but struggle with the global structure.
2
FILTR
Extracts the topological information contained in 3D models.
3
DONUT
Offers topological variety to train and evaluate models on topology.
Project Page
filtr-topology.github.io
Poster #2