Why open-world detectors fail to spot truly novel objects

Open-world object detectors can localize known classes but struggle to spot genuinely novel objects for future learning—their unknown predictions are mostly false alarms rather than true unknowns. The bottleneck isn't missing information; it's a one-dimensional objectness score that discards useful signal from the detector's internal representations. DualMem, a post-hoc filter using frozen SigLIP features and k-nearest-neighbor matching, reduces background false positives by 56.6% while preserving detection of true novel objects. It requires only a small labeled calibration set of future objects and lets users tune the precision–recall trade-off via Neyman-Pearson calibration.