[Paper Review] Predictive Inequity in Object Detection

Benjamin Wilson, Judy Hoffman+1 more|arXiv (Cornell University)|Feb 21, 2019

Advanced Neural Network ApplicationsComputer Science21 references154 citations

TL;DR

The paper examines whether state-of-the-art pedestrian detectors perform unequally across Fitzpatrick skin types, annotating BDD100K with skin tone labels and analyzing predictive disparity across models, data sources, and training losses.

ABSTRACT

In this work, we investigate whether state-of-the-art object detection systems have equitable predictive performance on pedestrians with different skin tones. This work is motivated by many recent examples of ML and vision systems displaying higher error rates for certain demographic groups than others. We annotate an existing large scale dataset which contains pedestrians, BDD100K, with Fitzpatrick skin tones in ranges [1-3] or [4-6]. We then provide an in-depth comparative analysis of performance between these two skin tone groupings, finding that neither time of day nor occlusion explain this behavior, suggesting this disparity is not merely the result of pedestrians in the 4-6 range appearing in more difficult scenes for detection. We investigate to what extent time of day, occlusion, and reweighting the supervised loss during training affect this predictive bias.

Motivation & Objective

• Measure predictive inequity in pedestrian detection between Fitzpatrick skin types 1-3 (LS) and 4-6 (DS) on a large driving dataset.
• Assess whether time of day, occlusion, or loss prioritization explain observed disparities.
• Compare multiple model architectures and training data sources to evaluate robustness of the inequity.
• Investigate simple remediation via loss weighting to reduce predictive disparity.

Proposed method

• Annotate BDD100K pedestrians with Fitzpatrick skin types using Mechanical Turk to create LS and DS groups.
• Use AP, AP50, and AP75 metrics on the BDD100K validation set to quantify disparity between LS and DS across models.
• Evaluate multiple architectures (Faster R-CNN, Mask R-CNN) and backbones with weights from MS COCO or BDD100K training.
• Analyze sources of inequity by isolating occluded vs unoccluded pedestrians and daytime vs nighttime images.
• Test loss-weighting schemes in the detector’s classification loss to reduce LS/DS gaps.

Experimental results

Research questions

• RQ1 Do object detectors exhibit higher predictive accuracy for LS (1-3) than DS (4-6) pedestrians on BDD100K?
• RQ2 Are disparities attributable to occlusion, time of day, or loss function bias toward the majority group?
• RQ3 Does the choice of training data (MS COCO vs BDD100K) or model architecture amplify or mitigate predictive inequity?
• RQ4 Can reweighting the supervised loss reduce the LS/DS performance gap without harming overall performance?

Key findings

• Detectors consistently show higher AP for LS than DS, with the largest gap at AP75 (tight localization).
• Disparity persists across models, data sources (MS COCO vs BDD100K), and architectures (Faster R-CNN, Mask R-CNN).
• Occlusion removal improves both groups but does not eliminate the LS/DS gap.
• Time of day alone does not explain the disparity; daytime results show LS>DS, while nighttime results are inconclusive due to small DS sample.
• Reweighting DS loss in training can reduce the DS gap in some settings, indicating loss prioritization contributes to predictive inequity.