T-Rex-Omni: Integrating Negative Visual Prompt in Generic Object Detection

The introduction of T-Rex-Omni marks a significant evolution in object detection methodologies, transitioning from traditional closed-set systems to more flexible open-set paradigms. Current open-set detectors often struggle with visually similar distractors due to their reliance on positive prompts. T-Rex-Omni addresses this challenge by incorporating negative visual prompts, allowing for a more nuanced understanding of object differentiation. The framework features a unified visual prompt encoder and a training-free Negating Negative Computing module, which dynamically suppresses negative responses during probability calculations. This innovative approach has demonstrated remarkable zero-shot detection performance, significantly narrowing the gap between visual-prompted and text-prompted methods. With a performance metric of 51.2 AP_r on the LVIS-minival dataset, T-Rex-Omni not only enhances detection accuracy but also supports flexible deployment in both positive-only and joint posi…