FILTR: Extracting Topological Features from Pretrained 3D Models
Researchers have developed FILTR, a novel framework designed to extract topological features from pretrained 3D models. This approach adapts a transformer decoder to generate persistence diagrams, which summarize a shape's multiscale structure, directly from frozen encoders. While existing 3D encoders show limited global topological signal, FILTR effectively utilizes their outputs to approximate these diagrams, enabling data-driven extraction from raw point clouds. AI
IMPACT Enables data-driven extraction of topological features from 3D point clouds, potentially improving shape analysis and understanding in computer vision.