Image similarity has been extensively studied in computer vision. In recently
years, machine-learned models have shown their ability to encode more semantics
than traditional multivariate metrics. However, in labelling similarity,
assigning a numerical score to a pair of images is less intuitive than
determining if an image A is closer to a reference image R than another image
B. In this work, we present a novel approach for building an image similarity
model based on labelled data in the form of A:R vs B:R. We address the
challenges of sparse sampling in the image space (R, A, B) and biases in the
models trained with context-based data by using an ensemble model. In
particular, we employed two ML techniques to construct such an ensemble model,
namely dimensionality reduction and MLP regressors. Our testing results show
that the ensemble model constructed performs ~5% better than the best
individual context-sensitive models. They also performed better than the model
trained with mixed imagery data as well as existing similarity models, e.g.,
CLIP and DINO. This work demonstrate that context-based labelling and model
training can be effective when an appropriate ensemble approach is used to
alleviate the limitation due to sparse sampling.

基于标注数据的 A:R vs B:R 方法构建了图像相似度模型，并采用集成模型处理稀疏采样和基于上下文数据的偏差，测试结果表明，该集成模型优于单个上下文敏感模型、混合图像数据和现有相似度模型，证明基于上下文标注和模型训练在适当的集成方法下可以有效克服稀疏采样的限制。

使用一组上下文敏感模型的图像相似度

Image Similarity using An Ensemble of Context-Sensitive Models

We examine the problem of obtaining fair outcomes for individuals who choose
to share optional information with machine-learned models and those who do not
consent and keep their data undisclosed. We find that these non-consenting
users receive significantly lower prediction outcomes than justified by their
provided information alone. This observation gives rise to the overlooked
problem of how to ensure that users, who protect their personal data, are not
penalized. While statistical fairness notions focus on fair outcomes between
advantaged and disadvantaged groups, these fairness notions fail to protect the
non-consenting users. To address this problem, we formalize protection
requirements for models which (i) allow users to benefit from sharing optional
information and (ii) do not penalize them if they keep their data undisclosed.
We offer the first solution to this problem by proposing the notion of Optional
Feature Fairness (OFF), which we prove to be loss-optimal under our protection
requirements (i) and (ii). To learn OFF-compliant models, we devise a
model-agnostic data augmentation strategy with finite sample convergence
guarantees. Finally, we extensively analyze OFF on a variety of challenging
real-world tasks, models, and data sets with multiple optional features.

本文研究了在个人选择共享可选信息和选择不共享个人数据的情况下，如何确保公正的结果。为解决此问题，提出 Optional Feature Fairness (OFF) 的概念和数据增强策略，并在多个挑战性实际任务中进行了广泛分析。