Introduction
While modern computer vision models have achieved remarkable proficiency in identifying the semantic content of images—recognizing objects, scenes, and actions—they often remain "time-blind." Most state-of-the-art models are trained on contemporary datasets like ImageNet or LAION, which lack the nuanced historical variations necessary to distinguish between different decades. For historians, archivists, and forensic analysts, the ability to automatically date millions of undated historical photographs represents a significant breakthrough.
Researchers from the University of Amsterdam recently investigated this challenge using the OpenCLIP model. Their study, "Blind Dates," explores whether AI can accurately date historical photographs from 1950 to 1999 and identifies the specific visual cues—or "temporal markers"—that allow a model to navigate the 20th century.
Methodology
The research team employed OpenCLIP, an open-source implementation of OpenAI’s Contrastive Language-Image Pre-training (CLIP), and structured their experiment into three distinct phases:
- Zero-shot Classification: The model was prompted with the text "a photograph from the year $x$" (ranging from 1950 to 2000) to evaluate its inherent, pre-trained sense of time.
- Fine-tuning (Linear Probing): To move beyond the limitations of text prompts, the researchers trained a logistic classifier on top of OpenCLIP’s image embeddings. This forced the model to rely solely on visual features to predict the year.
- Content Analysis: Using Detectron2, the team identified specific objects within the photos (such as people, vehicles, and animals). They then applied Bayesian regression to determine how the presence of these objects influenced the model's dating accuracy.
The Dataset: The De Boer Archive
The study utilized the De Boer Scene Detection dataset, a collection of 39,866 grayscale historical press photographs from the De Boer newspaper agency. Unlike iconic historical images that might be present in a model's training data, these are "mundane" press photos—sports, landscapes, and street scenes—providing a more rigorous test for temporal cues. To investigate the impact of photographic medium on AI perception, the researchers also created a colorized version of the dataset using DeOldify.
Key Insights and Results
The "Grayscale Bias"
In the zero-shot phase, the model exhibited a significant bias toward the past, frequently predicting dates much earlier than the actual year of the photograph. The researchers discovered that OpenCLIP inherently associates "grayscale" with "old." Interestingly, when the photos were colorized, the zero-shot accuracy improved, and the bias shifted toward the future. This suggests that the model's initial "sense of history" is heavily influenced by the materiality of the image rather than its content.
Fine-tuning and Materiality
Once the model was fine-tuned, the performance gap between grayscale and colorized images virtually disappeared. The fine-tuned model achieved a Mean Absolute Error (MAE) of 6.65 years on grayscale images, a substantial improvement over the zero-shot MAE of 15.8 years. This indicates that with proper training, a model can learn to ignore the medium (B&W vs. Color) and focus on the actual "message"—the visual content—to determine the date.
Humans as Clocks
One of the most compelling findings was the role of human subjects as temporal markers. The presence of people in a photograph significantly reduced the dating error, bringing the MAE down from 7.2 years to 5.5 years in the fine-tuned model. This is likely because humans carry dense, time-specific information through fashion, hairstyles, and accessories that evolve rapidly decade-to-decade.
Vehicles vs. Nature
Similarly, modes of transport like cars and buses served as reliable indicators of time due to recognizable design evolutions. Conversely, natural elements like birds or horses increased the model's error. These subjects lack time-specific visual signatures, making them "temporally neutral" and harder for the AI to place on a timeline.
Conclusion and Future Work
The study demonstrates that while "out-of-the-box" AI models struggle with historical temporality, fine-tuning allows them to surpass human-level performance (which typically sits at an MAE of roughly 11 years). The research highlights that the AI's ability to date an image is a combination of recognizing material cues and identifying specific semantic markers like fashion and technology.
Future work may involve expanding the time range to include the 19th century and the digital era, as well as utilizing more diverse geographical datasets. Additionally, the researchers suggest that moving toward granular object detection—identifying a specific 1960s car model rather than just a "car"—could further refine the accuracy of automated historical dating.
Summary updated