Source-linked AI summary
A Step Toward More Inclusive People Annotations for Fairness
Candice Schumann, Susanna Ricco, Utsav Prabhu, Vittorio Ferrari, Caroline Pantofaru
TL;DR
Open Images’ original person annotations are non-exhaustive, and its subclass-labeling procedure can affect subgroup labels and detector training. The paper introduces MIAP, with more complete person boxes and perceived attributes, and analyzes differences between the original and new annotations. The annotation sets vary in scope and coverage, so the authors describe them as useful in different situations or as complementary information.
Problem
The paper examines how Open Images’ non-exhaustive person annotations and subclass-labeling procedure affect fairness-relevant labels and detector training.
Source: §1 · p. 1, §3.3 · p. 5Method
Results
The original and MIAP annotations vary in scope and coverage, making them useful in different situations or as complementary information.
Source: §7 · p. 9Takeaways & Limitations
The authors say their findings can inform use of the current dataset and future annotation efforts, and that exhaustive person boxes support fairness-related subgroup analysis.
Source: §6 · p. 8, §7 · p. 9The relabeling leaves parts of the original pipeline unaddressed, may propagate or amplify existing workflow biases through image selection, and does not improve age-range imbalance.
Source: §6.1 · p. 8
Abstract
from arXiv · showhide
The Open Images Dataset contains approximately 9 million images and is a widely accepted dataset for computer vision research. As is common practice for large datasets, the annotations are not exhaustive, with bounding boxes and attribute labels for only a subset of the classes in each image. In this paper, we present a new set of annotations on a subset of the Open Images dataset called the MIAP (More Inclusive Annotations for People) subset, containing bounding boxes and attributes for all of the people visible in those images. The attributes and labeling methodology for the MIAP subset were designed to enable research into model fairness. In addition, we analyze the original annotation methodology for the person class and its subclasses, discussing the resulting patterns in order to inform future annotation efforts. By considering both the original and exhaustive annotation sets, researchers can also now study how systematic patterns in training annotations affect modeling.
1 INTRODUCTION
Open Images’ person-subclass labeling can leave person boxes incomplete and affect subgroup-label accuracy. MIAP introduces more complete person boxes and perceived age-range and gender-presentation attributes for fairness analysis, while the paper also specifies limits on intended use.
1 INTRODUCTION
- The paper introduces MIAP in Open Images Extended to improve boxes for the five original person-subtree classes using a protocol designed for fairness analysis.The authors hope it encourages Open Images researchers to incorporate fairness analysis.Source: §1 · p. 1
- The original annotation procedure affected subgroup-label accuracy and left person-superclass boxes incomplete in some images.Open Images has bounding boxes for only 600 of its more than 19,000 image-level classes, and its annotations are not exhaustive.Source: §1 · pp. 1–2
- MIAP asks annotators to box all people when an image has a positive label for person, man, woman, boy, or girl, then add perceived age-range and gender-presentation attributes.The procedure is designed to address boxes missing through interactions among scene content, image classifiers, and human annotators.Source: §1 · p. 2
1.1 Summary of Contributions
- The paper’s contributions include analysis of the original annotation pipeline, an updated procedure, comparative annotation analysis, and public release of the new annotations.Source: §1.1 · p. 2
1.2 Intended Use
- The age-range and gender-presentation labels reflect visual assessment by a third party, not individuals’ self-identified gender or actual age.The authors do not support training or deploying age- or gender-presentation classifiers from these annotations, judging risks outside fairness research to outweigh potential benefits.Source: §1.2 · p. 2
2 RELATED WORK
Related work links dataset construction choices to bias in model performance and describes approaches to fairness evaluation and mitigation. MIAP differs from prior people-attribute datasets through its annotation approach and its coverage of complex, multi-person scenes.
2.1 Identifying and reducing bias in datasets
- Internet-image datasets can overrepresent some regions, and classifiers trained on these distributions have been more likely to misclassify examples from the developing world.Open Images and ImageNet contain disproportionate numbers of images from North America and Europe [25], with cited errors involving person- and object-related classes.Source: §2.1 · p. 2
- The Crowdsourced Extension adds images from underrepresented countries but focuses on non-human image-level labels and blurs people’s faces, limiting its suitability for person-detector analysis.Source: §2.1 · p. 2
- ImageNet’s person-subtree update analyzes label definitions and class representation, whereas this work focuses on Open Images person boxes and adds attribute labels.The authors hope the datasets can be used together in future research.Source: §2.1 · p. 2
2.2 Identifying and reducing bias in trained models
- Fairness evaluation often partitions results by attributes, transparency standards advocate group-disaggregated metrics, and some mitigation strategies require attribute labels.Prior people-image datasets include fairness benchmarks and datasets curated for face recognition or attribute-classifier training [18] [32].Source: §2.2 · pp. 2–3
- MIAP avoids learned attribute classifiers and annotates multiple people per image, including people whose faces are not clearly visible.Many comparable fairness datasets target face attributes or recognition and contain a single face of interest per image.Source: §2.2 · p. 3
- BDD100k annotates Fitzpatrick skin type for sufficiently large pedestrians and supports person-detector fairness evaluation, but is limited to self-driving; MIAP does not address skin tone.MIAP annotations also support fairness analysis across a broader set of scene-understanding tasks.Source: §2.2 · p. 3
3 ANALYSIS OF THE ORIGINAL OPEN IMAGES
Open Images' intended partial-label pipeline can still leave people unboxed because class-hierarchy choices interact with image-level labels, language, and scene context. The analysis identifies potential fairness risks from these structured omissions, while noting that their effects on model performance remain insufficiently understood.
3.1 The Annotation Pipeline
- Open Images annotations are intentionally partial: image-level labels precede boxes for the most-specific positive classes, leaving some objects and unknown classes unboxed.The protocol is standard practice, but the paper warns that partial labeling can have unintended consequences when annotating people.Source: §3.1 · p. 3
3.2 Annotation Pipeline Mixing with Societal Norms
- Dropping the man, woman, boy, and girl subclasses can discard many boxes, while remapping them to person does not guarantee complete people coverage.After remapping, a person is represented only if the superclass was proposed without a positive subclass, or the relevant subclass was proposed and agreed upon by both annotators.Source: §3.2 · pp. 3–4
- Uneven image-level labels plus subclass-preferred boxing can omit people: positive woman and suit labels miss men in suits, and the reverse labeling pattern can also occur.The paper says subtle omissions can persist even when all five person-related classes are positive.Source: §3.2 · p. 4
- Even with all person-related labels positive, annotation decisions can miss people who do not fit gender-stereotyped roles.The paper presents this as a potential omission, depending on annotators' specific decisions.Source: §3.2 · p. 4
Age and language.
- Missed boxes were more likely for images labeled girl or woman than man or boy, which the authors relate to differences in how those terms are used.Examples include a negative girl label without a woman replacement and a positive girl label that annotators applied to only some people.Source: Age and language · p. 5
Potential correlation with image context.
- Scene content affects proposed labels and thus box coverage; the paper gives wedding and military examples where stereotype-consistent people receive boxes while others are omitted.The text specifies a missing box for a woman in uniform and describes the wedding example as missing the man.Source: Potential correlation with image context · p. 5
3.3 Potential risks of direct use of non-exhaustive annotations
- Non-exhaustive person boxes can reinforce context-specific detection gaps for subpopulations, with consequences ranging from inconvenient search to life-safety failures.The authors note that models may imitate the annotation patterns, while the severity depends on the use case.Source: §3.3 · p. 5
- Because validation and test annotations are also non-exhaustive, disparate model performance may be difficult to detect, and the effects of structured annotation patterns remain unknown.The authors say a detailed study of these questions is future work supported by MIAP.Source: §3.3 · p. 5
4 MORE INCLUSIVE ANNOTATIONS FOR PEOPLE
MIAP expands person annotations by asking annotators to box all people in selected images and assign perceived age-range and gender-presentation attributes. The protocol treats these as image-based perceptions, not a person’s identity or actual age.
4 MORE INCLUSIVE ANNOTATIONS FOR PEOPLE
- Image context and gender norms sometimes dictate which image-level labels are flagged.Source: Figure 7 · p. 6
- Annotators were asked to box all people in selected images and provide two attributes for each person.The relabeling effort sampled 100,000 images containing at least one of the five person classes in the positive image-level set.Source: §4 · pp. 5–6
Perceived Gender Presentation.
- Gender presentation is coded as predominantly feminine, predominantly masculine, or unknown from visual cues—not identity; guidance warns against stereotypes and allows unknown when evidence is insufficient.Annotators were reminded that norms vary across cultures and over time, and could use either presentation label even when a face was not visible.Source: Perceived Gender Presentation · p. 6
Perceived Age Range.
- Annotators assign perceived age range as young, middle, older, or unknown, preferring the older category when uncertain at a boundary.Age range and gender presentation could be labeled independently, including when one attribute was unknown.Source: Perceived Age Range · p. 6
5 ANALYSIS
The analysis compares original Open Images annotations with MIAP to examine missing person boxes and attribute patterns. It finds frequent added boxes and systematic differences across labels and dataset splits.
Raw counts.
- Approximately 30% of the 100,000 images had additional boxes added during relabeling.Table 1 reports the distribution of selected images and counts images where MIAP added boxes absent from the original annotations.Source: §5 · p. 6
- The original and MIAP box counts are compared in aggregate and by attribute, with original attributes inferred from class labels.Source: §5 · p. 6, Table 2 · p. 7
- Unknown labels increased substantially in training but only slightly in validation and test, while gender-presentation labels remained majority predominantly masculine and middle was the most common age-range label.The authors attribute the split difference in unknown labels to the original annotation procedures; Table 2 compares box counts across splits and totals, assigning original attributes from class labels.Source: §5 · p. 7, Table 2 · p. 7
Statistical analysis.
- The analysis uses nPMI to compare co-occurrence of labels in images with missing boxes against images without missing boxes.The comparison is expressed as the difference between nPMI(Missing, label) and nPMI(Not Missing, label).Source: §5 · p. 7
- Unknown age-range and gender-presentation labels were more associated with missing boxes, while predominantly feminine labels showed a larger association than predominantly masculine labels.Age-range labels showed the least skew between images with and without missing boxes; Table 3 orders nPMI differences by significance.Source: §5 · p. 7, Table 3 · p. 8
Empirical visualization.
- Figure 8 shows four examples where MIAP adds boxes for people missing from the original annotations.The figure marks original boxes in magenta and new MIAP boxes in yellow.Source: §5 · p. 7, Figure 8 · p. 8
6 DISCUSSION
MIAP’s exhaustive person boxes support fairness analyses that partial annotations do not address effectively, including metrics stratified by perceived age range and gender presentation. The discussion also sets important boundaries: these presentation labels are not actual age or gender identity, and the relabeling retains limitations in data representation and annotation scope.
6 DISCUSSION
- Exhaustive person boxes support disaggregating or stratifying fairness metrics by perceived age range and gender presentation.The authors hope such analyses will eventually lead to less biased models, classifiers, and workflows that work equitably for all.Source: §6 · p. 8
- Perceived gender-presentation and age-range labels are for evaluation, not measures of gender identity or actual age.Boxes labeled young receive unknown gender presentation in the released dataset.Source: §6 · p. 8
6.1 Limitations
- The Flickr images underlying Open Images and this relabeled subset may not represent model uses, while geographic sampling and English-language construction introduce cultural bias.The cited work notes geographic sampling and English as the primary language for dataset construction and taxonomy definition.Source: §6.1 · p. 8
- The relabeling leaves pipeline stages unaddressed, retains masculine-presentation imbalance, does not improve age-range imbalance, and excludes other person-related labels and relationship-triplet reannotation.Image selection required a person-subclass label, which may have propagated or amplified biases in the original labeling workflow; triplets remain only for originally annotated boxes.Source: §6.1 · p. 8
- Using societal concepts of perceived gender for fairness analysis trades off against perpetuating binary gender definitions; future work could examine objective attributes such as makeup or facial hair.Source: §6.1 · p. 8
7 CONCLUSION
The paper presents MIAP, a new annotation procedure and annotations for an Open Images subset, with person boxes and perceived age-range and gender-presentation attributes designed for fairness analysis. Its analyses show that original and new annotations differ in scope and coverage, so they can serve different situations or complement one another.
7 CONCLUSION
- Dataset construction, annotation procedures, fairness analysis, and modeling under noisy ground truth continue to evolve.Source: §7 · p. 9
- MIAP provides a new annotation procedure and annotations for an Open Images subset, adding person boxes and perceived age-range and gender-presentation labels designed for fairness analysis.Source: §7 · p. 9
- The original and MIAP annotation pipelines produce different patterns, and their annotations are useful in different situations or as complementary information.The analysis underscores the importance of understanding annotation design decisions and their consequences.Source: §7 · p. 9
- Researchers can use MIAP’s exhaustive person annotations for fairness analysis and can expect more robust fairness evaluations.Source: §7 · p. 9
IMAGE CITATIONS
The paper’s images are licensed under Creative Commons By 2.0 from Open Images.
IMAGE CITATIONS
- Images used in the paper are licensed under Creative Commons By 2.0 from Open Images.Source: IMAGE CITATIONS · p. 9