8000 GitHub - ClaudiaCuttano/SANSA: Official repository for the paper "SANSA: Unleashing the Hidden Semantics in SAM2 for Few-Shot Segmentation."
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

Official repository for the paper "SANSA: Unleashing the Hidden Semantics in SAM2 for Few-Shot Segmentation."

Notifications You must be signed in to change notification settings

ClaudiaCuttano/SANSA

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

4 Commits
 
 
 
 
 
 

Repository files navigation

SANSA: Unleashing the Hidden Semantics in SAM2 for Few-Shot Segmentation

📄 arXiv Preprint
Claudia Cuttano, Gabriele Trivigno, Giuseppe Averta, Carlo Masone

Welcome to the official repository for SANSA, our paper: "SANSA: Unleashing the Hidden Semantics in SAM2 for Few-Shot Segmentation".

sansa_in_action.mp4

SANSA supports a wide range of prompts — including points, boxes, scribbles, and masks.
Given one or more reference images, it segments semantically corresponding objects in the target image.
The entire process runs in a unified pipeline, with no prompt-specific adjustments or external models.

🛠️ Coming Soon

  • 🔓 Open-source code and trained models
  • 💻 Try SANSA directly on your own data.

🌟 Why SANSA?

Beneath SAM2 tracking architecture lies a surprisingly rich semantic feature space. SANSA unveils this hidden structure and repurposes SAM2 into a powerful few-shot segmenter.

🎯 First solution to fully leverage SAM2 for few-shot segmentation — no external feature encoders, no prompt engineering.
🖱️ Prompt anything — points, boxes, scribbles, or masks.
3–5× faster, 4–5× smaller than prior methods.
🏆 State-of-the-art generalization to novel classes in few-shot segmentation benchmarks.


❓ Why Does It Work?

We extract SAM2 features from object instances across diverse images and visualize their distribution using the first three principal components from PCA.

While zero-shot features from SAM2 lack clear semantic structure, after adapting features with SANSA, we observe the emergence of well-defined semantic clusters: semantically similar instances group together, forming coherent clusters despite strong intra-class variation in visual appearance.

PCA Semantic Clusters

⚠️ These are not training classes: SANSA learns from base categories and generalizes to unseen ones by reorganizing the feature space for semantic alignment.


🔜 Code & Models

Code and pretrained models will be released soon.
Stay tuned for updates!

About

Official repository for the paper "SANSA: Unleashing the Hidden Semantics in SAM2 for Few-Shot Segmentation."

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published
0