pseudo-camera-10k dataset

Contents

This dataset contains 10k free images from world class photographers. The images have been resized using Lanczos antialiasing, with their smaller edge shifted to 1024px.

The aim of this dataset is a highly variable but high quality and high resolution set of images containing difficult concepts, with about half of the images being numbered group shots and family portraits with the number of subjects labeled.

No images were upsampled in this dataset. The results of downsampling were manually validated to ensure limited artifacting.

The images have been captioned with CogVLM.

Hardware

These images were captioned on a single A100-80G at a rate of roughly 3.8 seconds per image.

Use cases

These images may be used to train models.

Limitations

Natural photos can be highly noisy. This will be learnt by the model; that may be a desired trait.

The captions are not 100% accurate, but they are highly descriptive and close to 100%.

The images do not reflect the multitude of diversity that much larger sets would.