From Decoupled Heads to Self-Correcting Inference: SCOPE for Training-Free Open-Vocabulary Segmentation
Abstract
Training-free open-vocabulary semantic segmentation (OVSS) has progressed rapidly with foundation models, yet most pipelines still treat inference as a one-shot dense semantic readout, remaining static even when auxiliary refinement or external model composition is applied. We observe that SAM3 already natively exposes decoupled predictions with complementary semantic, instance-level, and presence-aware signals that can be reinterpreted as internal evidence for test-time self-correction. Based on this observation, we propose SCOPE, a fully training-free framework that reformulates OVSS as a self-correcting inference process through three tightly coupled stages: a Self-Diagnostic stage that identifies unreliable queries and triggers conditional re-inference, a Self-Correcting stage that reconciles semantic and instance predictions through asymmetric cross-head refinement, and a Self-Stabilizing stage that suppresses view-sensitive variance via aligned prediction fusion. SCOPE operates entirely within the native SAM3 inference stack without parameter updates or auxiliary backbones, and establishes a new state-of-the-art among training-free methods on eight standard benchmarks. Code and model are available at https://github.com/akuan1234/SCOPE.