Revisiting Out-of-Distribution Detection in Real-Time Object Detection: From Benchmark Pitfalls to a New Mitigation...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models struggle with out-of-distribution (OoD) inputs, leading to overconfident predictions on incorrect objects.
• Existing methods for OoD detection in object detection offer limited improvements by focusing on post-hoc scoring adjustments.

Purpose of the Study:

• To address critical overlooked issues in OoD detection for object detection models.
• To propose a novel training-time mitigation strategy to enhance model robustness against OoD inputs.

Main Methods:

• Analyzed widely used evaluation benchmarks, identifying significant data quality issues distorting performance metrics.
• Introduced a training-time paradigm involving fine-tuning detectors with synthesized OoD datasets that mimic in-distribution objects.
• Developed a method to suppress objectness for OoD objects, creating a more defensive decision boundary.

Main Results:

• Revealed up to 13% of objects in OoD test sets were mislabeled, impacting existing method evaluations.
• Achieved a 91% reduction in hallucination error for a YOLO model on the BDD-100K dataset.
• Demonstrated generalization across various object detection architectures (YOLO, Faster R-CNN, RT-DETR) and few-shot adaptation.

Conclusions:

• Current OoD detection benchmarks have fundamental flaws that inflate performance claims.
• A training-time mitigation approach, independent of external detectors, effectively reduces OoD-induced hallucination.
• The proposed methodology offers a principled and broadly applicable solution for improving object detector reliability.