DualMem: Bypassing the Objectness Bottleneck for Calibrated Unknown-Stream Filtering in Open-World Object Detection
Researchers have developed DualMem, a novel post-hoc filter designed to improve open-world object detection systems. This method addresses the issue of polluted unknown prediction streams in current detectors, where background false positives are common. DualMem utilizes frozen SigLIP features and a calibrated likelihood ratio test with positive and negative memory banks to effectively filter out unwanted proposals, significantly reducing false unknowns while preserving the detection of known objects. AI
IMPACT Enhances open-world object detection by reducing false positives, potentially improving systems that need to identify novel objects.