Title: LFA: Layer Feature Attention for Run-Time Introspection of 2D Object Detectors in Automated Driving

Abstract

While reliable object detection is a cornerstone of automated driving, even the most advanced detectors are prone to errors that can jeopardize safety. To mitigate these risks, introspection techniques that forecast potential detector failures allow for safer system deployment by activating fallback protocols or notifying human operators. However, current methods predominantly depend on hand-crafted statistics or features from the final layer alone, thereby neglecting valuable data from earlier stages that represent varying degrees of visual abstraction.

In this work, we introduce Layer Feature Attention (LFA), a computationally efficient introspection framework that employs an attention mechanism to synthesize features from multiple layers of the backbone network. Our central premise is that detection inaccuracies appear distinctively across the feature hierarchy: high-level layers provide the semantic context necessary for scene comprehension, whereas low-level layers retain the fine-grained details crucial for identifying small or partially occluded objects. By learning layer-specific importance weights in an end-to-end manner, LFA not only enhances the accuracy of error prediction but also offers interpretable insights into which feature levels are most predictive of failure. Comprehensive evaluations on the KITTI and BDD100K datasets reveal that LFA sets a new state-of-the-art for introspection performance, surpassing single-layer baseline approaches across various detector architectures.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC