All computer vision models are wrong/useful: reflections on automatic detection of gender in historical photographs

Engineering notes

Key topics: autonomous driving, prediction. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

This article examines the challenges of using computer vision models to detect gender in historical photographs. Drawing on a failed attempt to empirically map gendered spaces in 1930s Swedish photo archives, the study instead turns to reflect on the biases and limitations of popular models such as Grounding DINO, ViLT-VQA, and Llama. Grounded in the field of distant viewing and critical data studies, the article explores how these models – often shaped by contemporary training data and commercial priorities – struggle to process visual material outside our contemporary time period. Using a manually annotated validation set of 1,500 photographs from the cultural heritage platform DigitaltMuseum, the study demonstrates how gender predictions indicate anachronistic assumptions, misrecognitions, and confirmation biases. Yet, rather than framing these models as flawed tools in need of correction, the article turns their limitations into productive provocations for visual analysis, exposing algorithmic logics of automated vision, which in turn reveal biases in how humans interpret images. Along with recent scholars, the article advocates for a critical humanities approach, treating computer vision models not just as instruments of analysis but as objects of inquiry – models that both reveal and reshape the visual categories through which we interpret the past.

Links and sources

• PDF from original source

Comments