Reliable Evidence Learning for Small Targets in Unified Context-Dependent Concept Segmentation

Unified context-dependent concept segmentation aims to segment heterogeneous visual concepts with a shared model conditioned on support examples. Small targets challenge this paradigm because their sparse visual evidence is easily overwhelmed by background structures, while locally strong responses may still be inconsistent with the specified concept. We present RELIC, a framework for reliable evidence learning that distributes concept conditioning across representation adaptation, evidence verification, and final prediction. Specifically, Size- and Concept-aware Small-target Evidence Verification (SC-SEV) matches query features against foreground and background support descriptor groups and combines relative foreground–background evidence with an absolute foreground criterion. The verification criteria adapt to target occupancy and support–query agreement, allowing weak but concept-consistent responses to receive stronger emphasis. To refine the remaining errors, an Evidence-Guided Small-Target Mixture-of-Experts (EGST-MoE) combines query-level expert weighting with spatial gating for weak-response recovery, boundary refinement, fragment completion, and false-positive suppression. Without bells and whistles, RELIC achieves consistently strong performance across eight benchmarks spanning natural, industrial, remote-sensing, and medical scenarios, with particularly pronounced advantages in segmenting small and ultra-small targets.