Photo Dating by Facial Age Aggregation

TL;DR

This work tackles photo dating by leveraging temporal signals contained in faces within images. It proposes a probabilistic framework that combines age-based posteriors, identity-driven face likelihoods, and career-based temporal priors to infer the capture year, aggregating evidence across multiple recognizable individuals. To enable rigorous evaluation, it introduces CSFD-1.6M, a large-scale, multi-face dataset with identity birth years and image years, released to the community along with embeddings and age posteriors. Empirical results show that multi-face aggregation improves year estimates and can outperform a strong scene-based baseline, while also supporting age estimation research and demonstrating practical scalability and ethical considerations for real-world use.

Abstract

We introduce a novel method for Photo Dating which estimates the year a photograph was taken by leveraging information from the faces of people present in the image. To facilitate this research, we publicly release CSFD-1.6M, a new dataset containing over 1.6 million annotated faces, primarily from movie stills, with identity and birth year annotations. Uniquely, our dataset provides annotations for multiple individuals within a single image, enabling the study of multi-face information aggregation. We propose a probabilistic framework that formally combines visual evidence from modern face recognition and age estimation models, and career-based temporal priors to infer the photo capture year. Our experiments demonstrate that aggregating evidence from multiple faces consistently improves the performance and the approach significantly outperforms strong, scene-based baselines, particularly for images containing several identifiable individuals.

Figures (17)

• Figure 1: Overview of the proposed dating framework. Our method is demonstrated on an example image from CSFD-1.6M. (a) The input image with detected faces and their ground-truth identities noted. (b) For each detected face, we show the matched reference identity portrait and the corresponding year posterior, which is derived from the model's age estimate. (c) The temporal information from the recognized identities is aggregated into a joint movie prior for the capture year. (d) The final capture year posterior is obtained by combining the joint prior from (c) with the individual year posteriors from (b) using the model described in \ref{['eq:model_marginalization']}.
• Figure 2: Overview of the CSFD-1.6M dataset statistics.
• Figure 3: Performance comparison of temporal priors. MAE for the Full model using different priors, and MAE of the Naive and Scene baselines. The results highlight the benefit of informative priors. The realistic Decade Prior ($p_{\mathrm{D}}$) consistently outperforms the Naive baseline and the uninformative Uniform Prior ($p_{\mathrm{U}}$). The oracle priors ($p_{\mathrm{I}}$, $p_{\mathrm{M}}$, $p_{\mathrm{C}}$), which use statistics from the test set, are included to illustrate an upper bound on performance.
• Figure 4: Ablation of information sources. MAE of predictions made using only a temporal prior ($p_{\text{D}}$ or $p_{\text{I}}$), only the age posterior ($p_{\text{age}}$, equivalent to the Naive model), and the full model combining both ($p_{\text{D}}\,\&\, p_{\text{age}}$ or $p_{\text{I}}\,\&\, p_{\text{age}}$). The results demonstrate that both the temporal prior and the age posterior are informative signals. Combining them using the proposed aggregation consistently yields lower error than using either source in isolation.
• Figure 5: Impact of prior strength ($\lambda$) on model performance. MAE of the Full model using the combination prior $p_{\mathrm{C}}$, as $\lambda$ interpolates between a purely uniform prior ($\lambda=0$) and the full oracle image prior ($\lambda=1$). The results show that even weak prior information yields substantial performance gain over the baseline.