Uncovering Hidden Patterns: A Concept-Driven Approach for Out-of-Distribution Object Detection

Hang Yu

Abstract: Object detection models excel in controlled environments but often stumble in dynamic, real-world scenarios where unexpected entities appear alongside familiar ones. The conventional path—manipulating feature spaces with artificial outliers—provides some resilience, yet it falls short in forging a deep comprehension of novelty without exposure to true unknowns. This paper explores a fresh perspective: by embracing the intertwined nature of known and novel objects in everyday scenes, we can harness scene-wide patterns to spot anomalies. Enter SceneConcept, our innovative framework that reimagines OOD handling through a two-pronged strategy. It begins with a self-guided process to break down images into modular, concept-rich slots that capture essential relationships and sparse encodings for both routine and aberrant elements. From there, a flexible integration layer weaves these slots into the detector's workflow, instilling an intuitive sense of unfamiliarity. For practical use, we incorporate a scoring system tied to overall image dynamics, sharpening the line between what's expected and what's not. Through rigorous testing on established datasets, SceneConcept not only matches but surpasses prior benchmarks.