Neurosymbolic Logic Guides Weakly Supervised Segmentation
A new approach integrates differentiable fuzzy logic with deep segmentation models to improve weakly supervised semantic segmentation (WSSS). The method uses logical constraints to fine-tune the Segment Anything Model (SAM) under weak supervision, generating better pseudo-labels for training a second-stage prompt-free model. Evaluated on Pascal VOC 2012 and REFUGE2 datasets, the logic-guided fine-tuning outperforms heuristic prompt-based methods.
Key facts
- Weakly supervised semantic segmentation (WSSS) trains models from partial annotations like bounding boxes, scribbles, or image-level tags.
- Recent work uses foundation models such as SAM to generate pseudo-labels.
- Existing approaches depend on heuristic prompt choices and limited prior knowledge integration.
- The proposed method takes a neurosymbolic perspective, integrating differentiable fuzzy logic with deep segmentation models.
- Weak annotations and domain-specific priors are unified as continuous logical constraints.
- The refined foundation model produces improved pseudo-labels for training a second-stage prompt-free segmentation model.
- Experiments were conducted on Pascal VOC 2012 and REFUGE2 optic disc/cup segmentation datasets.
- The logic-guided fine-tuning shows improved performance over heuristic prompt-based methods.
Entities
Institutions
- arXiv