TL;DR
Open-world object detection systems struggle to identify unknown objects because their prediction streams contain mostly background false positives (46-71%) rather than actual novel objects (under 10%). DualMem, a post-hoc filtering method, uses frozen SigLIP image features and k-nearest-neighbor matching with Neyman-Pearson calibration to distinguish true unknowns from background noise.
✦ Why It Matters
Engineers can reduce false unknown detections by 45-66% in open-world systems without retraining detectors or sacrificing known-class accuracy.
Key Takeaways
How It Works
DualMem operates by using a small, annotated calibration set of future-task objects to perform a non-parametric likelihood ratio test in a frozen feature space. It maintains a positive memory for future-task objects and a negative memory to filter out background-like proposals, effectively distinguishing between true unknowns and false positives.
Related