DualMem: New OWOD Method Calibrates Unknown-Stream Filtering
There's a new method called DualMem for open-world object detection, and it aims to fix the problem of mixed-up unknown predictions. Studies show that in strong OWOD systems like PROB, OW-DETR, and HypOW, less than 10% of future-task positive unknowns show up in unknown predictions, while false positives from the background can hit between 46% and 71%. This issue stems from a bottleneck in the objectness head. For instance, in PROB Task 1, a linear probe on a 256-D decoder query has an AUROC of 0.908 for telling apart positive and negative unknowns, but the final objectness scalar drops to 0.642. Using a frozen SigLIP feature boosts this separability during filtering (AUROC = 0.871). DualMem provides a smart filtering approach to address this issue, and you can find the full research paper on arXiv with ID 2605.23634.
Key facts
- DualMem is a new method for open-world object detection.
- Unknown prediction streams in OWOD detectors are heavily polluted.
- Future-task positive unknowns make up less than 10% of unknown predictions.
- Background false positives account for 46-71% of unknown predictions.
- The problem is an information bottleneck at the objectness head.
- On PROB Task 1, a linear probe on 256-D decoder query achieves AUROC 0.908.
- The final one-dimensional objectness scalar drops to AUROC 0.642.
- A frozen SigLIP feature achieves AUROC 0.871 at filtering stage.
Entities
Institutions
- arXiv