C-GAP's iterative prompting boosts vision-language minority detection by 81%

Safety-critical perception systems must detect rare objects reliably, but existing long-tailed detection methods assume hundreds of classes and dense annotations—an impractical fit. Open-vocabulary detectors, which accept natural language queries at inference, make prompt quality a direct lever for performance. Rather than retraining or collecting more data, C-GAP exploits this property by iteratively refining prompts for frozen detectors.

C-GAP operates in two phases. First, it builds a composite caption baseline that fuses per-image scene descriptions with class-quantity context—outperforming prompts using either alone across multiple architectures and datasets. Second, an LLM refines each image's caption individually, with trials triaged into accept, tentative, or regenerate buckets based on minority-class AP@0.5 against a dynamic threshold. Refinement stops once sufficient gain is achieved, and no detector weights are updated. Experiments show C-GAP lifts minority-class AP by up to 53% across benchmarks; on COCO, relative improvement reaches 81% (17.69 → 32.09 AP@0.5). The composite baseline is critical: starting from scene-only or class-quantity-only prompts yields diminishing returns.