Fairness in Autonomous Driving: Towards Understanding Confounding Factors in Object Detection under Challenging Weather

TL;DR

This work tackles fairness in object detection for autonomous driving under challenging weather by evaluating a state-of-the-art transformer detector (DETR) on both real (FACET) and simulated (Carla) data. It introduces probability-based fairness metrics that extend beyond traditional AR and AP to include confidence-based measures like ATPC and AFPC, enabling a more nuanced audit of group disparities under varying lighting, skin tones, and distances. The study reveals that demographic attributes such as skin tone and body size interact with weather and scene composition to affect detection performance, with disparities that can shift under different evaluation setups, and highlights the importance of using both real data and high-fidelity simulations to understand these effects. The findings advocate for testbeds that audit fairness in AV perception pipelines and point to downstream control implications, proposing directions for future work on true negatives and false positives in safety-critical decision-making.

Abstract

The deployment of autonomous vehicles (AVs) is rapidly expanding to numerous cities. At the heart of AVs, the object detection module assumes a paramount role, directly influencing all downstream decision-making tasks by considering the presence of nearby pedestrians, vehicles, and more. Despite high accuracy of pedestrians detected on held-out datasets, the potential presence of algorithmic bias in such object detectors, particularly in challenging weather conditions, remains unclear. This study provides a comprehensive empirical analysis of fairness in detecting pedestrians in a state-of-the-art transformer-based object detector. In addition to classical metrics, we introduce novel probability-based metrics to measure various intricate properties of object detection. Leveraging the state-of-the-art FACET dataset and the Carla high-fidelity vehicle simulator, our analysis explores the effect of protected attributes such as gender, skin tone, and body size on object detection performance in varying environmental conditions such as ambient darkness and fog. Our quantitative analysis reveals how the previously overlooked yet intuitive factors, such as the distribution of demographic groups in the scene, the severity of weather, the pedestrians' proximity to the AV, among others, affect object detection performance. Our code is available at https://github.com/bimsarapathiraja/fair-AV.

Figures (14)

• Figure 1: Ground truth for a pedestrian (green), true positive for the pedestrian (purple) and the false positives (red) are shown. The IoU between green and purple bounding boxes is high and the IoUs between green and red bounding boxes are zero.
• Figure 2: Black bounding boxes in the bottom plots indicate ground truth bounding boxes (i.e., there is only one pedestrian). Green indicates true positives and red indicates false positives. The confidence score is shown only when it is > 0.5. With the high levels of fog, it is possible to get false positives with confidences as high as 0.8.
• Figure 3: Monk Skin Tone (MST) Monk_2019 scale where MST=1 is the lightest skin tone and MST=10 is the darkest skin tone
• Figure 4: Carla simulation sample image across fog intensities of 0%, 25%, 50%, 75%, and 100%. The visibility of the road incrementally reduces as the fog intensity increases.
• Figure 5: Histogram of the images and annotation with Monk Skin Tone scale for FACET dataset. Lighter skin tone annotations are more prominent in the dataset compared to the darker skin tone annotations.