# The Forchheim Image Database for Camera Identification in the Wild

**Authors:** Benjamin Hadwiger, Christian Riess  
**arXiv:** [2011.02241](https://arxiv.org/abs/2011.02241)  
**Date:** 2020-11-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2011.02241v1/

## Key points

- Importance of image provenance in criminal investigations and journalistic fact-checking
- Need for rigorous testing of algorithms for extracting information about source camera and distribution history of images in real-world scenarios
- Introduction of the Forchheim Image Database (FODB) with over 23,000 images from 27 different smartphone cameras across 143 scenes
- FODB allows clear separation of image content from forensic artifacts and provides images in six different qualities
- Demonstration of FODB's utility through evaluation of camera identification methods and presentation of three key findings:
- General-purpose algorithm outperforms dedicated forensic CNNs on clean and compressed images
- Classifiers show improved performance when faced with unknown post-processing after being augmented by artificial degradations
- FODB's ability to cleanly separate scene content from forensic traces sets rigorous boundary conditions for benchmarking algorithms

## Summary

In their paper titled "The Forchheim Image Database for Camera Identification in the Wild," authors Benjamin Hadwiger and Christian Riess discuss the importance of image provenance in criminal investigations and journalistic fact-checking. They highlight the need for rigorous testing of algorithms that extract information about the source camera and distribution history of an image in real-world scenarios. To address this gap, they introduce the Forchheim Image Database (FODB), which comprises over 23,000 images captured by 27 different smartphone cameras across 143 scenes. The FODB allows for clear separation of image content from forensic artifacts and provides each image in six different qualities. The authors demonstrate its utility by evaluating methods for camera identification and present three key findings. Firstly, they observe that a general-purpose algorithm outperforms dedicated forensic CNNs on both clean and compressed images. Secondly, classifiers show improved performance when faced with unknown post-processing after being augmented by artificial degradations. Lastly, FODB's ability to cleanly separate scene content from forensic traces sets rigorous boundary conditions for benchmarking algorithms in this domain. Overall, this study underscores the significance of accurate image provenance in investigative processes and showcases how FODB fills a critical gap in existing datasets by providing a comprehensive resource for testing camera identification methods under realistic conditions.