---
type: slide
---

### Foundation Model Assisted Weakly Supervised Semantic Segmentation

---

#### Methods

- **CLIP**: Used with frozen weights for image classification and seed segmentation tasks.
- **SAM**: Provides class-agnostic object masks, improving the precision of segmentation seeds.

---

#### Framework Overview


---

#### SAM-based seeding module

- Quasi-superpixel generation
- lower threshold $t_{m,whole}$ to retrieve more masks at the *whole* level
- prioritze the selection of the *whole* level
- apply filter

- Quasi-superpixel classification

- seed map generation

---

#### Quasi-superpixel generartion


---

- Quasi-superpixel classification
foerground: 
background: 
semantic class: 

- seed map generation

---

- Experimental result


---

---