Prior2Former
Evidential Modeling of Mask Transformers for Open-World Panoptic Segmentation
🏆 ICCV 2025 — Highlight
Overview
How can segmentation models reliably handle objects they have never seen before? In safety-critical applications like autonomous driving, this is not just an academic question — it is a matter of life and death.
Prior2Former (P2F) is the first evidential learning approach for segmentation vision transformers. By incorporating a Beta prior into the mask transformer architecture, P2F computes principled model uncertainty for pixel-wise binary mask assignments — enabling it to detect novel and out-of-distribution objects that other models simply miss.
Key Contributions
- 🧠 First evidential mask transformer — introduces Beta prior-based uncertainty estimation into the mask vision transformer architecture
- 🚫 No OOD data required — unlike competing methods, P2F needs no access to OOD samples or contrastive void-class training
- 🔄 Flexible application — can be applied to both anomaly instance segmentation and open-world panoptic segmentation
- 🏅 State-of-the-art results across Cityscapes, COCO, SegmentMeIfYouCan, and OoDIS benchmarks
Why It Matters
Most panoptic segmentation models fail silently on unknown objects. Prior2Former closes this gap by providing high-quality uncertainty estimates that flag novel categories — making it a critical building block for trustworthy autonomous perception systems.