RT-DETRv4: Painlessly Furthering Real-Time Object Detection with Vision Foundation Models
The RT-DETRv4 model family enhances real-time object detection by using Vision Foundation Models (VFMs) to improve feature quality without increasing speed or deployment cost. The method uses a Deep Semantic Injector (DSI) to transfer rich semantics from VFMs into the detector's features, stabilized by Gradient-guided Adaptive Modulation (GAM). This approach achieves state-of-the-art COCO scores across multiple scales while maintaining high inference speeds.