Curated by: Ryan Ramos, Vladan Stojnic, Giorgos Kordopatis-Zilos, Yuta Nakashima, Giorgos Tolias, Noa Garcia
Funded by: Dataset collection was supported by JSPS KAKENHI No. JP23H00497 and JP22K12091, JST CREST Grant No. JPMJCR20D3, and JST FOREST Grant No. JPMJFR216O.
License: CC BY 4.0

Dataset Sources

Repository: https://github.com/ryan-caesar-ramos/visual-encoder-traces
Paper: Processing and acquisition traces in visual encoders: What does CLIP know about your camera?
Summary thread: here

Uses

Direct Use

This dataset is intended to be used for probing visual encoders' abilities to encode acquisition labels into their embeddings.

Out-of-Scope Use

Due to potential correlations between image semantics and acquisition labels, we do not recommend using this dataset to evaluate acquisition label prediction without preprocessing (e.g. masking, as done in the paper this dataset accompanies). Additional, due to limitations described below, we do not endorse the use of this dataset for tasks where balanced data is extremely important (e.g. image generation).

Dataset Structure

Flickr ID: the identifier of the image on Flickr
url: URL to the Flickr image
owner: identifier of the account that uploaded the image to Flickr
Make: the manufacturer of the camera used to capture the image
Model: the camera model used to capture the image
Exposure: the amount of time that light that was allowed to enter the camera while capturing the image
Aperture: the size of the opening in the lens and the corresponding amount of light thus allowed to enter the camera while capturing the image
ISO Speed: the sensitivity to light of the camera sensor used to capture the image
Focal Length: the distance between the center of the camera's lens and the camera's sensor

If your goal is to use this dataset to recreate our results, note that there is no single train set. We create splits per attribute, derive new attributes, and manually intervene on some labels. To recreate our results, please use the JSON files under split_data/. You can use the following snipped to download them:

from huggingface_hub import snapshot_download

snapshot_download(
  repo_id="ryanramos/FlickrExif",
  allow_patterns="split_data/*.json",
  repo_type="dataset"
)

Dataset Creation

Curation Rationale

Many datasets to do not include Exif metadata, and those that include them to do not cover acquisition parameters such as camera model. We close this gap with FlickrExif.

Source Data

Data Collection and Processing

We used the FlickrAPI to search for 2,000-4,000 safe-for-work, permissively licensed images each month from January 2000 to August 2025. To prevent the dataset from being dominated by prolific photographers, we limit the number of images per photographer to 10 per month and year.

Who are the source data producers?

This data is collected from Flickr users who marked their images are safe-for-work and permissively licensed, according to Flickr API's search filters.

Personal and Sensitive Information

Flickr user IDs are included in this dataset which can be used to access their profile pages.

Bias, Risks, and Limitations

There may exist correlations between an image's semantic content and its acquisition labels. For example, nighttime photos may be associated with higher ISO values. This can potentially affect acquisition label prediction if models rely on semantic cues as shortcuts. Furthermore, as this dataset was created with a specific use case in mind, the only filtering or balancing used to create this dataset is the owner filtering process described in the in the Data Collection and Processing section. Any remaining harmful content or imbalance resulting from the API's search algorithm are left untouched.

Recommendations

In acquisition label prediction settings, we highly recommend users to scrub semantic content from images with techniques such as masking, as done in the paper this dataset accompanies. Furthermore, because of the additional limitations mentioned previously, we do not endorse the use of FlickrExif for use cases outside its intended one.

Citation

BibTeX:

@misc{ramos2025processingacquisitiontracesvisual,
      title={Processing and acquisition traces in visual encoders: What does CLIP know about your camera?}, 
      author={Ryan Ramos and Vladan Stojnić and Giorgos Kordopatis-Zilos and Yuta Nakashima and Giorgos Tolias and Noa Garcia},
      year={2025},
      eprint={2508.10637},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.10637}, 
}

APA:

Ryan Ramos, Vladan Stojnić, Giorgos Kordopatis-Zilos, Yuta Nakashima, Giorgos Tolias, & Noa Garcia. (2025). Processing and acquisition traces in visual encoders: What does CLIP know about your camera?.

Dataset Card Authors

Ryan Ramos

Dataset Card Contact

[email protected]

Downloads last month: 36